[PATCH 0/4] Add Panthor compute-shader test

[PATCH 0/4] Add Panthor compute-shader test

[Daniel Stone]

Hi,
This series adds a test to the Panthor suite to exercise a simple
compute shader. Whilst the existing group_submit test does validate that
we're able to execute CSF commands, it doesn't validate that we can use
this to actually run any shaders. This series should help as minimal
bringup and validation, inspired by a platform which was able to pass
group_submit just fine, but would immediately wedge the firmware when
trying to run a compute-shader test.

I don't plan to extend this much further - at least not with more
complex GPU command types or shader ISA, since that would require a lot
more divergent per-arch code to deal with changes in descriptor tables
and commands - but it does ensure that we can actually schedule and
execute jobs on a GPU.

Tested on v10 (RK3588), v13 (MT8196), and v14 (Lumex CSS FVP).

Cheers,
Daniel

[PATCH 1/4] panthor: Use all available GPU cores

[Daniel Stone]

There's no reason to limit ourselves to one arbitrarily-chosen core.

Signed-off-by: Daniel Stone <daniels@collabora.com>
---
 lib/igt_panthor.c | 29 +++++++----------------------
 lib/igt_panthor.h |  1 -
 2 files changed, 7 insertions(+), 23 deletions(-)

diff --git a/lib/igt_panthor.c b/lib/igt_panthor.c
index 11d8ec342..cd6b6b7fd 100644
--- a/lib/igt_panthor.c
+++ b/lib/igt_panthor.c
@@ -2,6 +2,7 @@
 // SPDX-FileCopyrightText: Copyright (C) 2025 Collabora Ltd.
 
 #include "drmtest.h"
+#include "igt_aux.h"
 #include "igt_panthor.h"
 #include "ioctl_wrappers.h"
 #include "panthor_drm.h"
@@ -69,22 +70,6 @@ void igt_panthor_group_submit(int fd, struct drm_panthor_group_submit *group_sub
 		do_ioctl(fd, DRM_IOCTL_PANTHOR_GROUP_SUBMIT, group_submit);
 }
 
-/**
- * igt_panthor_get_first_core:
- * @cores_present: bitmask of available cores
- *
- * Get a mask with only the first available core bit set.
- *
- * Returns: core mask with first available core, or 0 if no cores available
- */
-uint64_t igt_panthor_get_first_core(uint64_t cores_present)
-{
-	if (cores_present == 0)
-		return 0;
-
-	return 1ULL << (ffs(cores_present) - 1);
-}
-
 /**
  * igt_panthor_group_create_simple:
  * @fd: device file descriptor
@@ -109,13 +94,13 @@ uint32_t igt_panthor_group_create_simple(int fd, uint32_t vm_id, int err)
 	queues = (struct drm_panthor_obj_array)DRM_PANTHOR_OBJ_ARRAY(1, &queue);
 
 	group_create.queues = queues;
-	group_create.max_compute_cores = 1;
-	group_create.max_fragment_cores = 1;
-	group_create.max_tiler_cores = 1;
+	group_create.max_compute_cores = igt_hweight(gpu_info.shader_present);
+	group_create.max_fragment_cores = igt_hweight(gpu_info.shader_present);
+	group_create.max_tiler_cores = igt_hweight(gpu_info.tiler_present);
 	group_create.priority = PANTHOR_GROUP_PRIORITY_MEDIUM;
-	group_create.compute_core_mask = igt_panthor_get_first_core(gpu_info.shader_present);
-	group_create.fragment_core_mask = igt_panthor_get_first_core(gpu_info.shader_present);
-	group_create.tiler_core_mask = igt_panthor_get_first_core(gpu_info.tiler_present);
+	group_create.compute_core_mask = gpu_info.shader_present;
+	group_create.fragment_core_mask = gpu_info.shader_present;
+	group_create.tiler_core_mask = gpu_info.tiler_present;
 	group_create.vm_id = vm_id;
 
 	igt_panthor_group_create(fd, &group_create, err);
diff --git a/lib/igt_panthor.h b/lib/igt_panthor.h
index ba4a9bf1e..63c4f2bf4 100644
--- a/lib/igt_panthor.h
+++ b/lib/igt_panthor.h
@@ -36,7 +36,6 @@ void igt_panthor_group_submit_simple(int fd, uint32_t group_handle,
 				     uint32_t queue_index, uint64_t stream_addr,
 				     uint32_t stream_size, uint32_t syncobj_handle,
 				     int err);
-uint64_t igt_panthor_get_first_core(uint64_t cores_present);
 
 static inline void igt_panthor_vm_bind(int fd, uint32_t vm_id, uint32_t bo_handle,
 				       uint64_t va, uint64_t size, uint32_t flags, int err)
-- 
2.47.3

Re: [PATCH 1/4] panthor: Use all available GPU cores

[Kamil Konieczny]

Hi Daniel,
On 2026-10-15 at 14:05:46 +0100, Daniel Stone wrote:
> There's no reason to limit ourselves to one arbitrarily-chosen core.
> 
> Signed-off-by: Daniel Stone <daniels@collabora.com>
LGTM
Acked-by: Kamil Konieczny <kamil.konieczny@linux.intel.com>

Regards,
Kamil

> ---
>  lib/igt_panthor.c | 29 +++++++----------------------
>  lib/igt_panthor.h |  1 -
>  2 files changed, 7 insertions(+), 23 deletions(-)
> 
> diff --git a/lib/igt_panthor.c b/lib/igt_panthor.c
> index 11d8ec342..cd6b6b7fd 100644
> --- a/lib/igt_panthor.c
> +++ b/lib/igt_panthor.c
> @@ -2,6 +2,7 @@
>  // SPDX-FileCopyrightText: Copyright (C) 2025 Collabora Ltd.
>  
>  #include "drmtest.h"
> +#include "igt_aux.h"
>  #include "igt_panthor.h"
>  #include "ioctl_wrappers.h"
>  #include "panthor_drm.h"
> @@ -69,22 +70,6 @@ void igt_panthor_group_submit(int fd, struct drm_panthor_group_submit *group_sub
>  		do_ioctl(fd, DRM_IOCTL_PANTHOR_GROUP_SUBMIT, group_submit);
>  }
>  
> -/**
> - * igt_panthor_get_first_core:
> - * @cores_present: bitmask of available cores
> - *
> - * Get a mask with only the first available core bit set.
> - *
> - * Returns: core mask with first available core, or 0 if no cores available
> - */
> -uint64_t igt_panthor_get_first_core(uint64_t cores_present)
> -{
> -	if (cores_present == 0)
> -		return 0;
> -
> -	return 1ULL << (ffs(cores_present) - 1);
> -}
> -
>  /**
>   * igt_panthor_group_create_simple:
>   * @fd: device file descriptor
> @@ -109,13 +94,13 @@ uint32_t igt_panthor_group_create_simple(int fd, uint32_t vm_id, int err)
>  	queues = (struct drm_panthor_obj_array)DRM_PANTHOR_OBJ_ARRAY(1, &queue);
>  
>  	group_create.queues = queues;
> -	group_create.max_compute_cores = 1;
> -	group_create.max_fragment_cores = 1;
> -	group_create.max_tiler_cores = 1;
> +	group_create.max_compute_cores = igt_hweight(gpu_info.shader_present);
> +	group_create.max_fragment_cores = igt_hweight(gpu_info.shader_present);
> +	group_create.max_tiler_cores = igt_hweight(gpu_info.tiler_present);
>  	group_create.priority = PANTHOR_GROUP_PRIORITY_MEDIUM;
> -	group_create.compute_core_mask = igt_panthor_get_first_core(gpu_info.shader_present);
> -	group_create.fragment_core_mask = igt_panthor_get_first_core(gpu_info.shader_present);
> -	group_create.tiler_core_mask = igt_panthor_get_first_core(gpu_info.tiler_present);
> +	group_create.compute_core_mask = gpu_info.shader_present;
> +	group_create.fragment_core_mask = gpu_info.shader_present;
> +	group_create.tiler_core_mask = gpu_info.tiler_present;
>  	group_create.vm_id = vm_id;
>  
>  	igt_panthor_group_create(fd, &group_create, err);
> diff --git a/lib/igt_panthor.h b/lib/igt_panthor.h
> index ba4a9bf1e..63c4f2bf4 100644
> --- a/lib/igt_panthor.h
> +++ b/lib/igt_panthor.h
> @@ -36,7 +36,6 @@ void igt_panthor_group_submit_simple(int fd, uint32_t group_handle,
>  				     uint32_t queue_index, uint64_t stream_addr,
>  				     uint32_t stream_size, uint32_t syncobj_handle,
>  				     int err);
> -uint64_t igt_panthor_get_first_core(uint64_t cores_present);
>  
>  static inline void igt_panthor_vm_bind(int fd, uint32_t vm_id, uint32_t bo_handle,
>  				       uint64_t va, uint64_t size, uint32_t flags, int err)
> -- 
> 2.47.3
> 

Re: [PATCH 1/4] panthor: Use all available GPU cores

[Liviu Dudau]

On Thu, Oct 15, 2026 at 02:05:46PM +0100, Daniel Stone wrote:
> There's no reason to limit ourselves to one arbitrarily-chosen core.
> 
> Signed-off-by: Daniel Stone <daniels@collabora.com>

Reviewed-by: Liviu Dudau <liviu.dudau@arm.com>

Best regards,
Liviu

> ---
>  lib/igt_panthor.c | 29 +++++++----------------------
>  lib/igt_panthor.h |  1 -
>  2 files changed, 7 insertions(+), 23 deletions(-)
> 
> diff --git a/lib/igt_panthor.c b/lib/igt_panthor.c
> index 11d8ec342..cd6b6b7fd 100644
> --- a/lib/igt_panthor.c
> +++ b/lib/igt_panthor.c
> @@ -2,6 +2,7 @@
>  // SPDX-FileCopyrightText: Copyright (C) 2025 Collabora Ltd.
>  
>  #include "drmtest.h"
> +#include "igt_aux.h"
>  #include "igt_panthor.h"
>  #include "ioctl_wrappers.h"
>  #include "panthor_drm.h"
> @@ -69,22 +70,6 @@ void igt_panthor_group_submit(int fd, struct drm_panthor_group_submit *group_sub
>  		do_ioctl(fd, DRM_IOCTL_PANTHOR_GROUP_SUBMIT, group_submit);
>  }
>  
> -/**
> - * igt_panthor_get_first_core:
> - * @cores_present: bitmask of available cores
> - *
> - * Get a mask with only the first available core bit set.
> - *
> - * Returns: core mask with first available core, or 0 if no cores available
> - */
> -uint64_t igt_panthor_get_first_core(uint64_t cores_present)
> -{
> -	if (cores_present == 0)
> -		return 0;
> -
> -	return 1ULL << (ffs(cores_present) - 1);
> -}
> -
>  /**
>   * igt_panthor_group_create_simple:
>   * @fd: device file descriptor
> @@ -109,13 +94,13 @@ uint32_t igt_panthor_group_create_simple(int fd, uint32_t vm_id, int err)
>  	queues = (struct drm_panthor_obj_array)DRM_PANTHOR_OBJ_ARRAY(1, &queue);
>  
>  	group_create.queues = queues;
> -	group_create.max_compute_cores = 1;
> -	group_create.max_fragment_cores = 1;
> -	group_create.max_tiler_cores = 1;
> +	group_create.max_compute_cores = igt_hweight(gpu_info.shader_present);
> +	group_create.max_fragment_cores = igt_hweight(gpu_info.shader_present);
> +	group_create.max_tiler_cores = igt_hweight(gpu_info.tiler_present);
>  	group_create.priority = PANTHOR_GROUP_PRIORITY_MEDIUM;
> -	group_create.compute_core_mask = igt_panthor_get_first_core(gpu_info.shader_present);
> -	group_create.fragment_core_mask = igt_panthor_get_first_core(gpu_info.shader_present);
> -	group_create.tiler_core_mask = igt_panthor_get_first_core(gpu_info.tiler_present);
> +	group_create.compute_core_mask = gpu_info.shader_present;
> +	group_create.fragment_core_mask = gpu_info.shader_present;
> +	group_create.tiler_core_mask = gpu_info.tiler_present;
>  	group_create.vm_id = vm_id;
>  
>  	igt_panthor_group_create(fd, &group_create, err);
> diff --git a/lib/igt_panthor.h b/lib/igt_panthor.h
> index ba4a9bf1e..63c4f2bf4 100644
> --- a/lib/igt_panthor.h
> +++ b/lib/igt_panthor.h
> @@ -36,7 +36,6 @@ void igt_panthor_group_submit_simple(int fd, uint32_t group_handle,
>  				     uint32_t queue_index, uint64_t stream_addr,
>  				     uint32_t stream_size, uint32_t syncobj_handle,
>  				     int err);
> -uint64_t igt_panthor_get_first_core(uint64_t cores_present);
>  
>  static inline void igt_panthor_vm_bind(int fd, uint32_t vm_id, uint32_t bo_handle,
>  				       uint64_t va, uint64_t size, uint32_t flags, int err)
> -- 
> 2.47.3
> 

-- 
====================
| I would like to |
| fix the world,  |
| but they're not |
| giving me the   |
 \ source code!  /
  ---------------
    ¯\_(ツ)_/¯

[PATCH 2/4] panthor: Add igt_panthor_get_arch() helper

[Daniel Stone]

This will identify the major architecture revision.

Signed-off-by: Daniel Stone <daniels@collabora.com>
---
 lib/igt_panthor.c | 17 +++++++++++++++++
 lib/igt_panthor.h |  1 +
 2 files changed, 18 insertions(+)

diff --git a/lib/igt_panthor.c b/lib/igt_panthor.c
index cd6b6b7fd..f5c6859e2 100644
--- a/lib/igt_panthor.c
+++ b/lib/igt_panthor.c
@@ -370,3 +370,20 @@ void igt_panthor_free_bo(int fd, struct panthor_bo *bo)
 
 	gem_close(fd, bo->handle);
 }
+
+/**
+ * igt_panthor_get_arch:
+ * @fd: panthor device file descriptor
+ *
+ * Get the architecture major version for this device.
+ *
+ * Returns: The Mali architecture major version
+ */
+int igt_panthor_get_arch(int fd)
+{
+	struct drm_panthor_gpu_info gpu_info = { 0 };
+
+	igt_panthor_query(fd, DRM_PANTHOR_DEV_QUERY_GPU_INFO,
+			  &gpu_info, sizeof(gpu_info), 0);
+	return DRM_PANTHOR_ARCH_MAJOR(gpu_info.gpu_id);
+}
diff --git a/lib/igt_panthor.h b/lib/igt_panthor.h
index 63c4f2bf4..3056ebd75 100644
--- a/lib/igt_panthor.h
+++ b/lib/igt_panthor.h
@@ -36,6 +36,7 @@ void igt_panthor_group_submit_simple(int fd, uint32_t group_handle,
 				     uint32_t queue_index, uint64_t stream_addr,
 				     uint32_t stream_size, uint32_t syncobj_handle,
 				     int err);
+int igt_panthor_get_arch(int fd);
 
 static inline void igt_panthor_vm_bind(int fd, uint32_t vm_id, uint32_t bo_handle,
 				       uint64_t va, uint64_t size, uint32_t flags, int err)
-- 
2.47.3

Re: [PATCH 2/4] panthor: Add igt_panthor_get_arch() helper

[Kamil Konieczny]

Hi Daniel,
On 2026-10-15 at 14:05:47 +0100, Daniel Stone wrote:

could you change subject and add a lib prefix? like

[PATCH 2/4] lib/igt_panthor: Add igt_panthor_get_arch() helper

This apply to 1...3 patches from series.

> This will identify the major architecture revision.
> 
> Signed-off-by: Daniel Stone <daniels@collabora.com>
LGTM
Acked-by: Kamil Konieczny <kamil.konieczny@linux.intel.com>

Regards,
Kamil

> ---
>  lib/igt_panthor.c | 17 +++++++++++++++++
>  lib/igt_panthor.h |  1 +
>  2 files changed, 18 insertions(+)
> 
> diff --git a/lib/igt_panthor.c b/lib/igt_panthor.c
> index cd6b6b7fd..f5c6859e2 100644
> --- a/lib/igt_panthor.c
> +++ b/lib/igt_panthor.c
> @@ -370,3 +370,20 @@ void igt_panthor_free_bo(int fd, struct panthor_bo *bo)
>  
>  	gem_close(fd, bo->handle);
>  }
> +
> +/**
> + * igt_panthor_get_arch:
> + * @fd: panthor device file descriptor
> + *
> + * Get the architecture major version for this device.
> + *
> + * Returns: The Mali architecture major version
> + */
> +int igt_panthor_get_arch(int fd)
> +{
> +	struct drm_panthor_gpu_info gpu_info = { 0 };
> +
> +	igt_panthor_query(fd, DRM_PANTHOR_DEV_QUERY_GPU_INFO,
> +			  &gpu_info, sizeof(gpu_info), 0);
> +	return DRM_PANTHOR_ARCH_MAJOR(gpu_info.gpu_id);
> +}
> diff --git a/lib/igt_panthor.h b/lib/igt_panthor.h
> index 63c4f2bf4..3056ebd75 100644
> --- a/lib/igt_panthor.h
> +++ b/lib/igt_panthor.h
> @@ -36,6 +36,7 @@ void igt_panthor_group_submit_simple(int fd, uint32_t group_handle,
>  				     uint32_t queue_index, uint64_t stream_addr,
>  				     uint32_t stream_size, uint32_t syncobj_handle,
>  				     int err);
> +int igt_panthor_get_arch(int fd);
>  
>  static inline void igt_panthor_vm_bind(int fd, uint32_t vm_id, uint32_t bo_handle,
>  				       uint64_t va, uint64_t size, uint32_t flags, int err)
> -- 
> 2.47.3
> 

Re: [PATCH 2/4] panthor: Add igt_panthor_get_arch() helper

[Liviu Dudau]

On Thu, Oct 15, 2026 at 02:05:47PM +0100, Daniel Stone wrote:
> This will identify the major architecture revision.
> 
> Signed-off-by: Daniel Stone <daniels@collabora.com>

Reviewed-by: Liviu Dudau <liviu.dudau@arm.com>

Best regards,
Liviu

> ---
>  lib/igt_panthor.c | 17 +++++++++++++++++
>  lib/igt_panthor.h |  1 +
>  2 files changed, 18 insertions(+)
> 
> diff --git a/lib/igt_panthor.c b/lib/igt_panthor.c
> index cd6b6b7fd..f5c6859e2 100644
> --- a/lib/igt_panthor.c
> +++ b/lib/igt_panthor.c
> @@ -370,3 +370,20 @@ void igt_panthor_free_bo(int fd, struct panthor_bo *bo)
>  
>  	gem_close(fd, bo->handle);
>  }
> +
> +/**
> + * igt_panthor_get_arch:
> + * @fd: panthor device file descriptor
> + *
> + * Get the architecture major version for this device.
> + *
> + * Returns: The Mali architecture major version
> + */
> +int igt_panthor_get_arch(int fd)
> +{
> +	struct drm_panthor_gpu_info gpu_info = { 0 };
> +
> +	igt_panthor_query(fd, DRM_PANTHOR_DEV_QUERY_GPU_INFO,
> +			  &gpu_info, sizeof(gpu_info), 0);
> +	return DRM_PANTHOR_ARCH_MAJOR(gpu_info.gpu_id);
> +}
> diff --git a/lib/igt_panthor.h b/lib/igt_panthor.h
> index 63c4f2bf4..3056ebd75 100644
> --- a/lib/igt_panthor.h
> +++ b/lib/igt_panthor.h
> @@ -36,6 +36,7 @@ void igt_panthor_group_submit_simple(int fd, uint32_t group_handle,
>  				     uint32_t queue_index, uint64_t stream_addr,
>  				     uint32_t stream_size, uint32_t syncobj_handle,
>  				     int err);
> +int igt_panthor_get_arch(int fd);
>  
>  static inline void igt_panthor_vm_bind(int fd, uint32_t vm_id, uint32_t bo_handle,
>  				       uint64_t va, uint64_t size, uint32_t flags, int err)
> -- 
> 2.47.3
> 

-- 
====================
| I would like to |
| fix the world,  |
| but they're not |
| giving me the   |
 \ source code!  /
  ---------------
    ¯\_(ツ)_/¯

[PATCH 3/4] panthor: Add more CSF instructions

[Daniel Stone]

Add a few more CSF instructions which we will need to execute shader
workloads.

Signed-off-by: Daniel Stone <daniels@collabora.com>
---
 lib/igt_panthor.h | 126 ++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 126 insertions(+)

diff --git a/lib/igt_panthor.h b/lib/igt_panthor.h
index 3056ebd75..dfeb0ac60 100644
--- a/lib/igt_panthor.h
+++ b/lib/igt_panthor.h
@@ -64,8 +64,13 @@ enum cs_opcode {
 	CS_OPCODE_MOVE48 = 1,
 	CS_OPCODE_MOVE32 = 2,
 	CS_OPCODE_WAIT = 3,
+	CS_OPCODE_RUN_COMPUTE = 4,
 	CS_OPCODE_STM = 21,
+	CS_OPCODE_SET_SB_ENTRY = 23,
+	CS_OPCODE_CALL = 32,
+	CS_OPCODE_REQ_RESOURCE = 34,
 	CS_OPCODE_FLUSH_CACHE = 36,
+	CS_OPCODE_ERROR_BARRIER = 47,
 };
 
 enum cs_flush_mode {
@@ -75,6 +80,12 @@ enum cs_flush_mode {
 	CS_FLUSH_MODE_CLEAN_AND_INVALIDATE = 3,
 };
 
+enum cs_task_axis {
+	CS_TASK_AXIS_X = 0,
+	CS_TASK_AXIS_Y = 1,
+	CS_TASK_AXIS_Z = 2,
+};
+
 /* There's no plan to support big endian in the UMD, so keep
  * things simple and don't bother supporting it.
  */
@@ -108,6 +119,19 @@ struct cs_instr {
 			uint64_t unused1: 23;
 			uint64_t opcode: 8;
 		} wait;
+		struct {
+			uint64_t task_increment: 14;
+			uint64_t task_axis: 2;
+			uint64_t unused0: 16;
+			uint64_t progress: 1;
+			uint64_t unused1: 7;
+			uint64_t srt_entry: 2;
+			uint64_t spd_entry: 2;
+			uint64_t tsd_entry: 2;
+			uint64_t fau_entry: 2;
+			uint64_t unused2: 8;
+			uint64_t opcode: 8;
+		} run_compute;
 		struct {
 			uint64_t offset: 16;
 			uint64_t mask: 16;
@@ -116,6 +140,27 @@ struct cs_instr {
 			uint64_t src: 8;
 			uint64_t opcode: 8;
 		} stm;
+		struct {
+			uint64_t ep_entry: 4;
+			uint64_t other_entry: 4;
+			uint64_t unused: 48;
+			uint64_t opcode: 8;
+		} set_sb_entry;
+		struct {
+			uint64_t unused0: 32;
+			uint64_t len: 8;
+			uint64_t addr: 8;
+			uint64_t unused1: 8;
+			uint64_t opcode: 8;
+		} call;
+		struct {
+			uint64_t compute: 1;
+			uint64_t fragment: 1;
+			uint64_t tiler: 1;
+			uint64_t idvs: 1;
+			uint64_t unused: 52;
+			uint64_t opcode: 8;
+		} req_resource;
 		struct {
 			uint64_t l2_mode: 4;
 			uint64_t lsc_mode: 4;
@@ -128,6 +173,10 @@ struct cs_instr {
 			uint64_t unused2: 4;
 			uint64_t opcode: 8;
 		} flush;
+		struct {
+			uint64_t unused: 56;
+			uint64_t opcode: 8;
+		} error_barrier;
 		uint64_t raw;
 	};
 };
@@ -186,6 +235,27 @@ cs_wait(uint16_t wait_mask, bool progress_inc)
 	return instr.raw;
 }
 
+static inline uint64_t
+cs_run_compute(uint16_t task_increment, enum cs_task_axis task_axis, bool progress,
+	       uint32_t srt_entry, uint32_t spd_entry, uint32_t tsd_entry,
+	       uint32_t fau_entry)
+{
+	struct cs_instr instr = {
+		.run_compute = {
+			.task_increment = task_increment,
+			.task_axis = task_axis,
+			.progress = progress,
+			.srt_entry = srt_entry,
+			.spd_entry = spd_entry,
+			.tsd_entry = tsd_entry,
+			.fau_entry = fau_entry,
+			.opcode = CS_OPCODE_RUN_COMPUTE,
+		},
+	};
+
+	return instr.raw;
+}
+
 static inline uint64_t
 cs_stm(uint8_t address, uint8_t src, uint16_t mask, int16_t offset)
 {
@@ -214,6 +284,50 @@ cs_stm64(uint8_t address, uint8_t src, int16_t offset)
 	return cs_stm(address, src, 0x3, offset);
 }
 
+static inline uint64_t
+cs_set_sb_entry(uint32_t ep_entry, uint32_t other_entry)
+{
+	struct cs_instr instr = {
+		.set_sb_entry = {
+			.ep_entry = ep_entry,
+			.other_entry = other_entry,
+			.opcode = CS_OPCODE_SET_SB_ENTRY,
+		},
+	};
+
+	return instr.raw;
+}
+
+static inline uint64_t
+cs_call(uint8_t addr, uint8_t len)
+{
+	struct cs_instr instr = {
+		.call = {
+			.addr = addr,
+			.len = len,
+			.opcode = CS_OPCODE_CALL,
+		},
+	};
+
+	return instr.raw;
+}
+
+static inline uint64_t
+cs_req_resource(bool compute, bool fragment, bool tiler, bool idvs)
+{
+	struct cs_instr instr = {
+		.req_resource = {
+			.compute = compute,
+			.fragment = fragment,
+			.tiler = tiler,
+			.idvs = idvs,
+			.opcode = CS_OPCODE_REQ_RESOURCE,
+		},
+	};
+
+	return instr.raw;
+}
+
 static inline uint64_t
 cs_flush(enum cs_flush_mode l2_mode,
 	 enum cs_flush_mode lsc_mode,
@@ -237,4 +351,16 @@ cs_flush(enum cs_flush_mode l2_mode,
 	return instr.raw;
 }
 
+static inline uint64_t
+cs_error_barrier(void)
+{
+	struct cs_instr instr = {
+		.error_barrier = {
+			.opcode = CS_OPCODE_ERROR_BARRIER,
+		},
+	};
+
+	return instr.raw;
+}
+
 #endif /* IGT_PANTHOR_H */
-- 
2.47.3

Re: [PATCH 3/4] panthor: Add more CSF instructions

[Kamil Konieczny]

Hi Daniel,
On 2026-10-15 at 14:05:48 +0100, Daniel Stone wrote:
> Add a few more CSF instructions which we will need to execute shader
> workloads.
> 
> Signed-off-by: Daniel Stone <daniels@collabora.com>
LGTM
Acked-by: Kamil Konieczny <kamil.konieczny@linux.intel.com>

Regards,
Kamil

> ---
>  lib/igt_panthor.h | 126 ++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 126 insertions(+)
> 
> diff --git a/lib/igt_panthor.h b/lib/igt_panthor.h
> index 3056ebd75..dfeb0ac60 100644
> --- a/lib/igt_panthor.h
> +++ b/lib/igt_panthor.h
> @@ -64,8 +64,13 @@ enum cs_opcode {
>  	CS_OPCODE_MOVE48 = 1,
>  	CS_OPCODE_MOVE32 = 2,
>  	CS_OPCODE_WAIT = 3,
> +	CS_OPCODE_RUN_COMPUTE = 4,
>  	CS_OPCODE_STM = 21,
> +	CS_OPCODE_SET_SB_ENTRY = 23,
> +	CS_OPCODE_CALL = 32,
> +	CS_OPCODE_REQ_RESOURCE = 34,
>  	CS_OPCODE_FLUSH_CACHE = 36,
> +	CS_OPCODE_ERROR_BARRIER = 47,
>  };
>  
>  enum cs_flush_mode {
> @@ -75,6 +80,12 @@ enum cs_flush_mode {
>  	CS_FLUSH_MODE_CLEAN_AND_INVALIDATE = 3,
>  };
>  
> +enum cs_task_axis {
> +	CS_TASK_AXIS_X = 0,
> +	CS_TASK_AXIS_Y = 1,
> +	CS_TASK_AXIS_Z = 2,
> +};
> +
>  /* There's no plan to support big endian in the UMD, so keep
>   * things simple and don't bother supporting it.
>   */
> @@ -108,6 +119,19 @@ struct cs_instr {
>  			uint64_t unused1: 23;
>  			uint64_t opcode: 8;
>  		} wait;
> +		struct {
> +			uint64_t task_increment: 14;
> +			uint64_t task_axis: 2;
> +			uint64_t unused0: 16;
> +			uint64_t progress: 1;
> +			uint64_t unused1: 7;
> +			uint64_t srt_entry: 2;
> +			uint64_t spd_entry: 2;
> +			uint64_t tsd_entry: 2;
> +			uint64_t fau_entry: 2;
> +			uint64_t unused2: 8;
> +			uint64_t opcode: 8;
> +		} run_compute;
>  		struct {
>  			uint64_t offset: 16;
>  			uint64_t mask: 16;
> @@ -116,6 +140,27 @@ struct cs_instr {
>  			uint64_t src: 8;
>  			uint64_t opcode: 8;
>  		} stm;
> +		struct {
> +			uint64_t ep_entry: 4;
> +			uint64_t other_entry: 4;
> +			uint64_t unused: 48;
> +			uint64_t opcode: 8;
> +		} set_sb_entry;
> +		struct {
> +			uint64_t unused0: 32;
> +			uint64_t len: 8;
> +			uint64_t addr: 8;
> +			uint64_t unused1: 8;
> +			uint64_t opcode: 8;
> +		} call;
> +		struct {
> +			uint64_t compute: 1;
> +			uint64_t fragment: 1;
> +			uint64_t tiler: 1;
> +			uint64_t idvs: 1;
> +			uint64_t unused: 52;
> +			uint64_t opcode: 8;
> +		} req_resource;
>  		struct {
>  			uint64_t l2_mode: 4;
>  			uint64_t lsc_mode: 4;
> @@ -128,6 +173,10 @@ struct cs_instr {
>  			uint64_t unused2: 4;
>  			uint64_t opcode: 8;
>  		} flush;
> +		struct {
> +			uint64_t unused: 56;
> +			uint64_t opcode: 8;
> +		} error_barrier;
>  		uint64_t raw;
>  	};
>  };
> @@ -186,6 +235,27 @@ cs_wait(uint16_t wait_mask, bool progress_inc)
>  	return instr.raw;
>  }
>  
> +static inline uint64_t
> +cs_run_compute(uint16_t task_increment, enum cs_task_axis task_axis, bool progress,
> +	       uint32_t srt_entry, uint32_t spd_entry, uint32_t tsd_entry,
> +	       uint32_t fau_entry)
> +{
> +	struct cs_instr instr = {
> +		.run_compute = {
> +			.task_increment = task_increment,
> +			.task_axis = task_axis,
> +			.progress = progress,
> +			.srt_entry = srt_entry,
> +			.spd_entry = spd_entry,
> +			.tsd_entry = tsd_entry,
> +			.fau_entry = fau_entry,
> +			.opcode = CS_OPCODE_RUN_COMPUTE,
> +		},
> +	};
> +
> +	return instr.raw;
> +}
> +
>  static inline uint64_t
>  cs_stm(uint8_t address, uint8_t src, uint16_t mask, int16_t offset)
>  {
> @@ -214,6 +284,50 @@ cs_stm64(uint8_t address, uint8_t src, int16_t offset)
>  	return cs_stm(address, src, 0x3, offset);
>  }
>  
> +static inline uint64_t
> +cs_set_sb_entry(uint32_t ep_entry, uint32_t other_entry)
> +{
> +	struct cs_instr instr = {
> +		.set_sb_entry = {
> +			.ep_entry = ep_entry,
> +			.other_entry = other_entry,
> +			.opcode = CS_OPCODE_SET_SB_ENTRY,
> +		},
> +	};
> +
> +	return instr.raw;
> +}
> +
> +static inline uint64_t
> +cs_call(uint8_t addr, uint8_t len)
> +{
> +	struct cs_instr instr = {
> +		.call = {
> +			.addr = addr,
> +			.len = len,
> +			.opcode = CS_OPCODE_CALL,
> +		},
> +	};
> +
> +	return instr.raw;
> +}
> +
> +static inline uint64_t
> +cs_req_resource(bool compute, bool fragment, bool tiler, bool idvs)
> +{
> +	struct cs_instr instr = {
> +		.req_resource = {
> +			.compute = compute,
> +			.fragment = fragment,
> +			.tiler = tiler,
> +			.idvs = idvs,
> +			.opcode = CS_OPCODE_REQ_RESOURCE,
> +		},
> +	};
> +
> +	return instr.raw;
> +}
> +
>  static inline uint64_t
>  cs_flush(enum cs_flush_mode l2_mode,
>  	 enum cs_flush_mode lsc_mode,
> @@ -237,4 +351,16 @@ cs_flush(enum cs_flush_mode l2_mode,
>  	return instr.raw;
>  }
>  
> +static inline uint64_t
> +cs_error_barrier(void)
> +{
> +	struct cs_instr instr = {
> +		.error_barrier = {
> +			.opcode = CS_OPCODE_ERROR_BARRIER,
> +		},
> +	};
> +
> +	return instr.raw;
> +}
> +
>  #endif /* IGT_PANTHOR_H */
> -- 
> 2.47.3
> 

Re: [PATCH 3/4] panthor: Add more CSF instructions

[Liviu Dudau]

On Thu, Oct 15, 2026 at 02:05:48PM +0100, Daniel Stone wrote:
> Add a few more CSF instructions which we will need to execute shader
> workloads.
> 
> Signed-off-by: Daniel Stone <daniels@collabora.com>

Reviewed-by: Liviu Dudau <liviu.dudau@arm.com>

Best regards,
Liviu

> ---
>  lib/igt_panthor.h | 126 ++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 126 insertions(+)
> 
> diff --git a/lib/igt_panthor.h b/lib/igt_panthor.h
> index 3056ebd75..dfeb0ac60 100644
> --- a/lib/igt_panthor.h
> +++ b/lib/igt_panthor.h
> @@ -64,8 +64,13 @@ enum cs_opcode {
>  	CS_OPCODE_MOVE48 = 1,
>  	CS_OPCODE_MOVE32 = 2,
>  	CS_OPCODE_WAIT = 3,
> +	CS_OPCODE_RUN_COMPUTE = 4,
>  	CS_OPCODE_STM = 21,
> +	CS_OPCODE_SET_SB_ENTRY = 23,
> +	CS_OPCODE_CALL = 32,
> +	CS_OPCODE_REQ_RESOURCE = 34,
>  	CS_OPCODE_FLUSH_CACHE = 36,
> +	CS_OPCODE_ERROR_BARRIER = 47,
>  };
>  
>  enum cs_flush_mode {
> @@ -75,6 +80,12 @@ enum cs_flush_mode {
>  	CS_FLUSH_MODE_CLEAN_AND_INVALIDATE = 3,
>  };
>  
> +enum cs_task_axis {
> +	CS_TASK_AXIS_X = 0,
> +	CS_TASK_AXIS_Y = 1,
> +	CS_TASK_AXIS_Z = 2,
> +};
> +
>  /* There's no plan to support big endian in the UMD, so keep
>   * things simple and don't bother supporting it.
>   */
> @@ -108,6 +119,19 @@ struct cs_instr {
>  			uint64_t unused1: 23;
>  			uint64_t opcode: 8;
>  		} wait;
> +		struct {
> +			uint64_t task_increment: 14;
> +			uint64_t task_axis: 2;
> +			uint64_t unused0: 16;
> +			uint64_t progress: 1;
> +			uint64_t unused1: 7;
> +			uint64_t srt_entry: 2;
> +			uint64_t spd_entry: 2;
> +			uint64_t tsd_entry: 2;
> +			uint64_t fau_entry: 2;
> +			uint64_t unused2: 8;
> +			uint64_t opcode: 8;
> +		} run_compute;
>  		struct {
>  			uint64_t offset: 16;
>  			uint64_t mask: 16;
> @@ -116,6 +140,27 @@ struct cs_instr {
>  			uint64_t src: 8;
>  			uint64_t opcode: 8;
>  		} stm;
> +		struct {
> +			uint64_t ep_entry: 4;
> +			uint64_t other_entry: 4;
> +			uint64_t unused: 48;
> +			uint64_t opcode: 8;
> +		} set_sb_entry;
> +		struct {
> +			uint64_t unused0: 32;
> +			uint64_t len: 8;
> +			uint64_t addr: 8;
> +			uint64_t unused1: 8;
> +			uint64_t opcode: 8;
> +		} call;
> +		struct {
> +			uint64_t compute: 1;
> +			uint64_t fragment: 1;
> +			uint64_t tiler: 1;
> +			uint64_t idvs: 1;
> +			uint64_t unused: 52;
> +			uint64_t opcode: 8;
> +		} req_resource;
>  		struct {
>  			uint64_t l2_mode: 4;
>  			uint64_t lsc_mode: 4;
> @@ -128,6 +173,10 @@ struct cs_instr {
>  			uint64_t unused2: 4;
>  			uint64_t opcode: 8;
>  		} flush;
> +		struct {
> +			uint64_t unused: 56;
> +			uint64_t opcode: 8;
> +		} error_barrier;
>  		uint64_t raw;
>  	};
>  };
> @@ -186,6 +235,27 @@ cs_wait(uint16_t wait_mask, bool progress_inc)
>  	return instr.raw;
>  }
>  
> +static inline uint64_t
> +cs_run_compute(uint16_t task_increment, enum cs_task_axis task_axis, bool progress,
> +	       uint32_t srt_entry, uint32_t spd_entry, uint32_t tsd_entry,
> +	       uint32_t fau_entry)
> +{
> +	struct cs_instr instr = {
> +		.run_compute = {
> +			.task_increment = task_increment,
> +			.task_axis = task_axis,
> +			.progress = progress,
> +			.srt_entry = srt_entry,
> +			.spd_entry = spd_entry,
> +			.tsd_entry = tsd_entry,
> +			.fau_entry = fau_entry,
> +			.opcode = CS_OPCODE_RUN_COMPUTE,
> +		},
> +	};
> +
> +	return instr.raw;
> +}
> +
>  static inline uint64_t
>  cs_stm(uint8_t address, uint8_t src, uint16_t mask, int16_t offset)
>  {
> @@ -214,6 +284,50 @@ cs_stm64(uint8_t address, uint8_t src, int16_t offset)
>  	return cs_stm(address, src, 0x3, offset);
>  }
>  
> +static inline uint64_t
> +cs_set_sb_entry(uint32_t ep_entry, uint32_t other_entry)
> +{
> +	struct cs_instr instr = {
> +		.set_sb_entry = {
> +			.ep_entry = ep_entry,
> +			.other_entry = other_entry,
> +			.opcode = CS_OPCODE_SET_SB_ENTRY,
> +		},
> +	};
> +
> +	return instr.raw;
> +}
> +
> +static inline uint64_t
> +cs_call(uint8_t addr, uint8_t len)
> +{
> +	struct cs_instr instr = {
> +		.call = {
> +			.addr = addr,
> +			.len = len,
> +			.opcode = CS_OPCODE_CALL,
> +		},
> +	};
> +
> +	return instr.raw;
> +}
> +
> +static inline uint64_t
> +cs_req_resource(bool compute, bool fragment, bool tiler, bool idvs)
> +{
> +	struct cs_instr instr = {
> +		.req_resource = {
> +			.compute = compute,
> +			.fragment = fragment,
> +			.tiler = tiler,
> +			.idvs = idvs,
> +			.opcode = CS_OPCODE_REQ_RESOURCE,
> +		},
> +	};
> +
> +	return instr.raw;
> +}
> +
>  static inline uint64_t
>  cs_flush(enum cs_flush_mode l2_mode,
>  	 enum cs_flush_mode lsc_mode,
> @@ -237,4 +351,16 @@ cs_flush(enum cs_flush_mode l2_mode,
>  	return instr.raw;
>  }
>  
> +static inline uint64_t
> +cs_error_barrier(void)
> +{
> +	struct cs_instr instr = {
> +		.error_barrier = {
> +			.opcode = CS_OPCODE_ERROR_BARRIER,
> +		},
> +	};
> +
> +	return instr.raw;
> +}
> +
>  #endif /* IGT_PANTHOR_H */
> -- 
> 2.47.3
> 

-- 
====================
| I would like to |
| fix the world,  |
| but they're not |
| giving me the   |
 \ source code!  /
  ---------------
    ¯\_(ツ)_/¯

[PATCH 4/4] panthor: Add simple compute-shader test

[Daniel Stone]

This test launches a compute shader which performs RMW on a memory
array, to check that we can actually launch jobs and see the results.

Signed-off-by: Daniel Stone <daniels@collabora.com>
---
 lib/igt_panthor.h             | 204 +++++++++++++++++++++++++++++
 tests/panthor/panthor_group.c | 240 ++++++++++++++++++++++++++++++++++
 2 files changed, 444 insertions(+)

diff --git a/lib/igt_panthor.h b/lib/igt_panthor.h
index dfeb0ac60..35882fad7 100644
--- a/lib/igt_panthor.h
+++ b/lib/igt_panthor.h
@@ -363,4 +363,208 @@ cs_error_barrier(void)
 	return instr.raw;
 }
 
+struct gpu_src_reg {
+	uint8_t reg;
+	bool last;
+	uint8_t rep;
+};
+
+struct gpu_dst_reg {
+	uint8_t reg;
+};
+
+enum gpu_flow_ctrl {
+	FLOW_NONE = 0,
+	FLOW_WAIT0 = 0x1,
+	FLOW_WAIT1 = 0x2,
+	FLOW_WAIT01 = 0x3,
+	FLOW_WAIT2 = 0x4,
+	FLOW_WAIT02 = 0x5,
+	FLOW_WAIT12 = 0x6,
+	FLOW_WAIT012 = 0x7,
+	FLOW_WAIT0126 = 0x8,
+	FLOW_WAIT = 0x9,
+	FLOW_RECONVERGE = 0xa,
+	FLOW_WAIT_RESOURCE = 0xc,
+	FLOW_DISCARD = 0xd,
+	FLOW_END = 0xf,
+};
+
+enum gpu_reg_width {
+	REG_WIDTH_LO16 = 0x1,
+	REG_WIDTH_HI16 = 0x2,
+	REG_WIDTH_32 = 0x3,
+};
+
+static inline struct gpu_dst_reg
+gpu_dst_reg(uint8_t reg)
+{
+	struct gpu_dst_reg ret = {
+		.reg = reg
+	};
+	return ret;
+}
+
+static inline struct gpu_src_reg
+gpu_src_zero(void)
+{
+	struct gpu_src_reg ret = {
+		.reg = 255, /* architecturally invalid */
+		.last = false,
+		.rep = (0x3 << 6),
+	};
+	return ret;
+}
+
+static inline struct gpu_src_reg
+gpu_src_reg(uint8_t reg)
+{
+	struct gpu_src_reg ret = {
+		.reg = reg,
+		.last = false,
+		.rep = reg,
+	};
+	return ret;
+}
+
+static inline struct gpu_src_reg
+gpu_src_reg_last(uint8_t reg)
+{
+	struct gpu_src_reg ret = {
+		.reg = reg,
+		.last = true,
+		.rep = (reg | (1 << 6)),
+	};
+	return ret;
+}
+
+#define ins(x, v, s) x |= ((uint64_t) v << s)
+
+static inline uint64_t
+gpu_mov_imm_i32(struct gpu_dst_reg dst, int32_t imm, enum gpu_flow_ctrl flow)
+{
+	uint64_t ret = 0;
+	ins(ret, gpu_src_zero().rep, 0);
+	ins(ret, imm, 8);
+	ins(ret, dst.reg, 40);
+	ins(ret, REG_WIDTH_32, 46);
+	ins(ret, 0x0, 48); /* opcode */
+	ins(ret, 0x11, 52); /* class */
+	ins(ret, 0x0, 57); /* FAU */
+	ins(ret, flow, 59);
+	return ret;
+}
+
+static inline uint64_t
+gpu_lshift_or_i32(struct gpu_dst_reg dst, struct gpu_src_reg src,
+		  struct gpu_src_reg shift, struct gpu_src_reg or,
+		  enum gpu_flow_ctrl flow)
+{
+	uint64_t ret = 0;
+	ins(ret, src.rep, 0);
+	ins(ret, shift.rep, 8);
+	ins(ret, or.rep, 16);
+	ins(ret, 0x1, 24); /* logicop = or */
+	ins(ret, 0x4, 26); /* lanes = b0 */
+	ins(ret, 0x0, 30); /* dst.not = false */
+	ins(ret, 0x1, 32); /* shift = left */
+	ins(ret, 0x0, 35); /* src2.not = false */
+	ins(ret, 0x0, 36); /* widen = false */
+	ins(ret, dst.reg, 40);
+	ins(ret, REG_WIDTH_32, 46);
+	ins(ret, 0x4, 48); /* opcode */
+	ins(ret, 0xb, 52); /* class */
+	ins(ret, 0x0, 57); /* FAU */
+	ins(ret, flow, 59);
+	return ret;
+}
+
+static inline uint64_t
+gpu_iadd_u32(struct gpu_dst_reg dst, struct gpu_src_reg left,
+	     struct gpu_src_reg right, enum gpu_flow_ctrl flow)
+{
+	uint64_t ret = 0;
+	ins(ret, left.rep, 0);
+	ins(ret, right.rep, 8);
+	ins(ret, 0x0, 16); /* opcode */
+	ins(ret, 0x0, 26); /* widen1 */
+	ins(ret, 0x0, 30); /* saturate */
+	ins(ret, 0x0, 36); /* widen0 */
+	ins(ret, dst.reg, 40);
+	ins(ret, REG_WIDTH_32, 46);
+	ins(ret, 0x0, 48); /* data type */
+	ins(ret, 0xa, 52); /* class */
+	ins(ret, 0x0, 57); /* FAU */
+	ins(ret, flow, 59);
+	return ret;
+}
+
+static inline uint64_t
+gpu_imul_u32(struct gpu_dst_reg dst, struct gpu_src_reg left,
+	     struct gpu_src_reg right, enum gpu_flow_ctrl flow)
+{
+	uint64_t ret = 0;
+	ins(ret, left.rep, 0);
+	ins(ret, right.rep, 8);
+	ins(ret, 0xa, 16); /* opcode */
+	ins(ret, 0x0, 26); /* widen1 */
+	ins(ret, 0x0, 30); /* saturate */
+	ins(ret, 0x0, 36); /* widen0 */
+	ins(ret, dst.reg, 40);
+	ins(ret, REG_WIDTH_32, 46);
+	ins(ret, 0x0, 48); /* data type */
+	ins(ret, 0xa, 52); /* class */
+	ins(ret, 0x0, 57); /* FAU */
+	ins(ret, flow, 59);
+	return ret;
+}
+
+static inline uint64_t
+gpu_load_i32(struct gpu_dst_reg dst, struct gpu_src_reg addr_lo,
+	     struct gpu_src_reg addr_hi, uint16_t offset,
+	     uint8_t sb_slot, enum gpu_flow_ctrl flow)
+{
+	uint64_t ret = 0;
+	igt_assert_eq(addr_hi.reg, addr_lo.reg + 1);
+	ins(ret, addr_lo.rep, 0);
+	ins(ret, offset, 8);
+	ins(ret, 0x0, 24); /* access mode */
+	ins(ret, REG_WIDTH_32, 27);
+	ins(ret, sb_slot, 30);
+	ins(ret, 0x1, 33); /* staging register count */
+	ins(ret, 0x0, 36); /* load 32 */
+	ins(ret, 0x1, 39); /* no sign extension */
+	ins(ret, dst.reg, 40); /* base staging register */
+	ins(ret, 0x2, 46); /* write staging registers */
+	ins(ret, 0x0, 48); /* opcode */
+	ins(ret, 0x6, 52); /* class */
+	ins(ret, 0x0, 57); /* FAU */
+	ins(ret, flow, 59);
+	return ret;
+}
+
+static inline uint64_t
+gpu_store_i32(struct gpu_src_reg addr_lo, struct gpu_src_reg addr_hi,
+	      struct gpu_src_reg val, uint16_t offset,
+	      uint8_t sb_slot, enum gpu_flow_ctrl flow)
+{
+	uint64_t ret = 0;
+	igt_assert_eq(addr_hi.reg, addr_lo.reg + 1);
+	ins(ret, addr_lo.rep, 0);
+	ins(ret, offset, 8);
+	ins(ret, 0x0, 24); /* access mode */
+	ins(ret, REG_WIDTH_32, 27);
+	ins(ret, sb_slot, 30);
+	ins(ret, 0x1, 33); /* staging register count */
+	ins(ret, val.reg, 40); /* base staging register */
+	ins(ret, 0x1, 46); /* read staging registers */
+	ins(ret, 0x1, 48); /* opcode */
+	ins(ret, 0x6, 52); /* class */
+	ins(ret, 0x0, 57); /* FAU */
+	ins(ret, flow, 59);
+	return ret;
+}
+
+#undef ins
+
 #endif /* IGT_PANTHOR_H */
diff --git a/tests/panthor/panthor_group.c b/tests/panthor/panthor_group.c
index c79f8d629..feca42482 100644
--- a/tests/panthor/panthor_group.c
+++ b/tests/panthor/panthor_group.c
@@ -10,6 +10,133 @@
 #include "igt_syncobj.h"
 #include "panthor_drm.h"
 
+/*
+ * Record a simple compute program:
+ * {
+ *     extern uint32_t shared_buf[num_invocations];
+ *     extern const uint32_t my_invocation_id;
+ *     shared_buf[my_invocation_id] *= my_invocation_id;
+ * }
+ */
+static void
+record_compute_prog(uint8_t *csf_cmdbuf, unsigned int *csf_cmdbuf_size,
+		    uint8_t *gpu_cmdbuf, unsigned int *gpu_cmdbuf_size,
+		    uint32_t srt_gpuva, uint32_t fau_gpuva, uint32_t spd_gpuva,
+		    uint32_t tsd_gpuva, uint32_t data_gpuva)
+{
+	const uint8_t reg_id_global_x = 60; /* preloaded */
+	const uint8_t reg_addr_lo = 0;
+	const uint8_t reg_addr_hi = 1;
+	const uint8_t reg_val = 2;
+	const uint8_t reg_offset = 3;
+	const uint8_t reg_shift = 4;
+	const uint8_t compute_sr_base = 0;
+	const uint8_t zero_reg = 40;
+
+	uint64_t gpu_instrs[] = {
+		/* MOV_IMM.i32 rADDRLO, 0x10000 */
+		gpu_mov_imm_i32(gpu_dst_reg(reg_addr_lo), data_gpuva, FLOW_NONE),
+		/* MOV_IMM.i32 rADDRHI, 0x0 */
+		gpu_mov_imm_i32(gpu_dst_reg(reg_addr_hi), 0x0, FLOW_NONE),
+		/* MOV_IMM.i32 rSHIFT, 0x2 */
+		gpu_mov_imm_i32(gpu_dst_reg(reg_shift), 0x2, FLOW_NONE),
+		/* LSHIFT_OR.i32 rOFFSET, rGLOBALID, rSHIFT^, 0x0 */
+		gpu_lshift_or_i32(gpu_dst_reg(reg_offset),
+				  gpu_src_reg(reg_id_global_x),
+				  gpu_src_reg_last(reg_shift),
+				  gpu_src_zero(),
+				  FLOW_NONE),
+		/* IADD.u32 rADDRLO, rADDRLO^, rOFFSET^ */
+		gpu_iadd_u32(gpu_dst_reg(reg_addr_lo),
+			     gpu_src_reg_last(reg_addr_lo),
+			     gpu_src_reg_last(reg_offset),
+			     FLOW_NONE),
+		/* LOAD.i32.slot0.wait0 @rVAL, [rADDRLO:rADDRHI], offset:0 */
+		gpu_load_i32(gpu_dst_reg(reg_val),
+			     gpu_src_reg(reg_addr_lo),
+			     gpu_src_reg(reg_addr_hi),
+			     0x0,
+			     0, FLOW_WAIT0),
+		/* IMUL_u32 rVAL, rVAL^, rGLOBALID^ */
+		gpu_imul_u32(gpu_dst_reg(reg_val),
+			     gpu_src_reg_last(reg_val),
+			     gpu_src_reg_last(reg_id_global_x),
+			     FLOW_NONE),
+		/* STORE.i32.slot1.end @rVAL, [rADDRLO^:rADDRHI^], offset:0 */
+		gpu_store_i32(gpu_src_reg_last(reg_addr_lo),
+			      gpu_src_reg_last(reg_addr_hi),
+			      gpu_src_reg(reg_val),
+			      0x0,
+			      1, FLOW_END),
+	};
+
+	uint64_t csf_instrs[] = {
+		/* REQ_RESOURCE.compute */
+		cs_req_resource(true, false, false, false),
+		/* MOV32 r40, 0 (always force cache flush/invalidate) */
+		cs_mov32(zero_reg, 0),
+		/* FLUSH_CACHE.clean_invalidate_l2.clean_invalidate_lsc.invalidate_other.sb0 */
+		cs_flush(CS_FLUSH_MODE_CLEAN_AND_INVALIDATE,
+			 CS_FLUSH_MODE_CLEAN_AND_INVALIDATE,
+			 CS_FLUSH_MODE_INVALIDATE,
+			 0, zero_reg, 1),
+		/* WAIT (flush) */
+		cs_wait(0xff, false),
+		/* ShaderResourceTable 0-3 pointers */
+		cs_mov48(compute_sr_base +  0, srt_gpuva),
+		cs_mov48(compute_sr_base +  2, 0x222222222222),
+		cs_mov48(compute_sr_base +  4, 0x333333333333),
+		cs_mov48(compute_sr_base +  6, 0x444444444444),
+		/* FastAccessUniform table 0-3 pointers */
+		cs_mov48(compute_sr_base +  8, fau_gpuva),
+		cs_mov48(compute_sr_base + 10, 0x666666666666),
+		cs_mov48(compute_sr_base + 12, 0x777777777777),
+		cs_mov48(compute_sr_base + 14, 0x888888888888),
+		/* ShaderProgramDescriptor 0-3 pointers */
+		cs_mov48(compute_sr_base + 16, spd_gpuva),
+		cs_mov48(compute_sr_base + 18, 0x999999999999),
+		cs_mov48(compute_sr_base + 20, 0xaaaaaaaaaaaa),
+		cs_mov48(compute_sr_base + 22, 0xbbbbbbbbbbbb),
+		/* ThreadStorageDescriptor 0-3 pointers */
+		cs_mov48(compute_sr_base + 24, tsd_gpuva),
+		cs_mov48(compute_sr_base + 26, 0xdddddddddddd),
+		cs_mov48(compute_sr_base + 28, 0xeeeeeeeeeeee),
+		cs_mov48(compute_sr_base + 30, 0xffffffffffff),
+		/* global ID offset = 0 */
+		cs_mov32(compute_sr_base + 32, 0),
+		/* workgroup size 64x1x1, allow merge */
+		cs_mov32(compute_sr_base + 33, (63 << 0) | (0 << 10) | (0 << 20) | (1 << 31)),
+		/* job offset (x/y/z) = { 0, 0, 0 } */
+		cs_mov32(compute_sr_base + 34, 0),
+		cs_mov32(compute_sr_base + 35, 0),
+		cs_mov32(compute_sr_base + 36, 0),
+		/* job size (x/y/z) = { 64, 1, 1 } */
+		cs_mov32(compute_sr_base + 37, 64),
+		cs_mov32(compute_sr_base + 38, 1),
+		cs_mov32(compute_sr_base + 39, 1),
+		/* SET_SB_ENTRY.async_ep1.other1 */
+		cs_set_sb_entry(1, 1),
+		/* RUN_COMPUTE.increment64.x_axis.no_increment.srt0.spd0.tsd0.fau0 */
+		cs_run_compute(64, CS_TASK_AXIS_X, 0, 0, 0, 0, 0),
+		/* WAIT (RUN_COMPUTE) */
+		cs_wait(0xff, false),
+		/* FLUSH_CACHE.clean_invalidate_l2.clean_invalidate_lsc.invalidate_other.sb0 */
+		cs_flush(CS_FLUSH_MODE_CLEAN_AND_INVALIDATE,
+			 CS_FLUSH_MODE_CLEAN_AND_INVALIDATE,
+			 CS_FLUSH_MODE_INVALIDATE,
+			 0, zero_reg, 1),
+		/* WAIT (flush) */
+		cs_wait(0xff, false),
+		/* ERROR_BARRIER */
+		cs_error_barrier(),
+	};
+
+	memcpy(gpu_cmdbuf, gpu_instrs, sizeof(gpu_instrs));
+	*gpu_cmdbuf_size = sizeof(gpu_instrs);
+	memcpy(csf_cmdbuf, csf_instrs, sizeof(csf_instrs));
+	*csf_cmdbuf_size = sizeof(csf_instrs);
+}
+
 static size_t
 issue_store_multiple(uint8_t *cs, uint64_t kernel_va, uint32_t constant)
 {
@@ -118,6 +245,119 @@ int igt_main() {
 		igt_panthor_free_bo(fd, &result_bo);
 	}
 
+	igt_describe("Submit a simple compute job with memory access.");
+	igt_subtest("simple_compute") {
+		uint32_t vm_id;
+		uint32_t group_handle;
+		struct panthor_bo csf_cmdbuf_bo = {};
+		struct panthor_bo gpu_cmdbuf_bo = {};
+		struct panthor_bo resource_table_bo = {};
+		struct panthor_bo data_bo = {};
+		uint64_t csf_cmdbuf_gpuva;
+		uint32_t csf_cmdbuf_size;
+		uint64_t gpu_cmdbuf_gpuva;
+		uint32_t gpu_cmdbuf_size;
+		uint32_t resource_table_gpuva;
+		uint32_t data_gpuva;
+		uint32_t syncobj_handle;
+		uint32_t srt_offset;
+		uint32_t fau_offset;
+		uint32_t spd_offset;
+		uint32_t *spd_ptr;
+		uint32_t spd_gpuva;
+		uint32_t tsd_offset;
+		uint32_t *tsd_ptr;
+		int arch = igt_panthor_get_arch(fd);
+
+		igt_require(arch < 15);
+
+		igt_panthor_vm_create(fd, &vm_id, 0);
+
+		/* CSF ISA command stream */
+		csf_cmdbuf_gpuva = 0x1000000;
+		igt_panthor_bo_create_mapped(fd, &csf_cmdbuf_bo, 4096, 0, 0);
+		igt_panthor_vm_bind(fd, vm_id, csf_cmdbuf_bo.handle, csf_cmdbuf_gpuva,
+				    csf_cmdbuf_bo.size, DRM_PANTHOR_VM_BIND_OP_TYPE_MAP, 0);
+		memset(csf_cmdbuf_bo.map, 0, csf_cmdbuf_bo.size);
+
+		/* GPU ISA command stream */
+		gpu_cmdbuf_gpuva = 0x2000000;
+		igt_panthor_bo_create_mapped(fd, &gpu_cmdbuf_bo, 4096, 0, 0);
+		igt_panthor_vm_bind(fd, vm_id, gpu_cmdbuf_bo.handle, gpu_cmdbuf_gpuva,
+				    gpu_cmdbuf_bo.size, DRM_PANTHOR_VM_BIND_OP_TYPE_MAP, 0);
+		memset(gpu_cmdbuf_bo.map, 0, gpu_cmdbuf_bo.size);
+
+		/* resource tables passed to RUN_COMPUTE */
+		resource_table_gpuva = 0x3000000;
+		igt_panthor_bo_create_mapped(fd, &resource_table_bo, 4096, 0, 0);
+		igt_panthor_vm_bind(fd, vm_id, resource_table_bo.handle, resource_table_gpuva,
+				    resource_table_bo.size, DRM_PANTHOR_VM_BIND_OP_TYPE_MAP, 0);
+		memset(resource_table_bo.map, 0, resource_table_bo.size);
+		/* ShaderResourceTable descriptor (zeroed) */
+		srt_offset = 0;
+		/* FastAccessUniform descriptor (zeroed) */
+		fau_offset = 128;
+		/* ShaderProgramDescriptor */
+		spd_offset = 256;
+		spd_gpuva = resource_table_gpuva + spd_offset;
+		if (arch == 14)
+			spd_gpuva |= 2; /* register allocation propagated into ptr */
+		spd_ptr = (uint32_t *) ((unsigned char *) resource_table_bo.map + spd_offset);
+		spd_ptr[0] = 0x8 | (0x1 << 4); /* compute shader */
+		spd_ptr[0] |= (0 << 16) | (0 << 17) | (0 << 19); /* no NaN suppress, FTZ, inf suppress */
+		spd_ptr[0] |= (0 << 29) | (2 << 30); /* no JUMP_EX, max 32 registers */
+		spd_ptr[1] |= (1 << 12); /* preload GlobalInvocationID.x into r60 */
+		spd_ptr[2] = gpu_cmdbuf_gpuva; /* initial PC */
+		/* ThreadStorageDescriptor */
+		tsd_offset = 512;
+		tsd_ptr = (uint32_t *) ((unsigned char *) resource_table_bo.map + tsd_offset);
+		tsd_ptr[1] = 0x1f; /* max wls_instances */
+
+		/* CPU<->GPU shared data */
+		data_gpuva = 0x4000000;
+		igt_panthor_bo_create_mapped(fd, &data_bo, 16384, 0, 0);
+		igt_panthor_vm_bind(fd, vm_id, data_bo.handle, data_gpuva,
+				    data_bo.size, DRM_PANTHOR_VM_BIND_OP_TYPE_MAP, 0);
+		for (unsigned int i = 0; i < 64 * 64; i++) {
+			uint32_t *x = data_bo.map + (i * sizeof(uint32_t));
+			*x = i;
+		}
+
+		record_compute_prog(csf_cmdbuf_bo.map, &csf_cmdbuf_size,
+				    gpu_cmdbuf_bo.map, &gpu_cmdbuf_size,
+				    resource_table_gpuva + srt_offset,
+				    resource_table_gpuva + fau_offset,
+				    spd_gpuva,
+				    resource_table_gpuva + tsd_offset,
+				    data_gpuva);
+
+		group_handle = igt_panthor_group_create_simple(fd, vm_id, 0);
+		igt_assert_neq(group_handle, 0);
+
+		syncobj_handle = syncobj_create(fd, 0);
+
+		igt_panthor_group_submit_simple(fd, group_handle, 0,
+						csf_cmdbuf_gpuva, csf_cmdbuf_size,
+						syncobj_handle, 0);
+		igt_assert(syncobj_wait(fd, &syncobj_handle, 1, INT64_MAX, 0, NULL));
+
+		for (unsigned int i = 0; i < 64 * 64; i++) {
+			uint32_t *x = data_bo.map + (i * sizeof(uint32_t));
+			igt_assert_eq(*x, i * i);
+		}
+
+		syncobj_destroy(fd, syncobj_handle);
+
+		igt_panthor_group_destroy(fd, group_handle, 0);
+
+		igt_panthor_vm_destroy(fd, vm_id, 0);
+
+		igt_panthor_free_bo(fd, &data_bo);
+		igt_panthor_free_bo(fd, &resource_table_bo);
+		igt_panthor_free_bo(fd, &gpu_cmdbuf_bo);
+		igt_panthor_free_bo(fd, &csf_cmdbuf_bo);
+	}
+
 	igt_fixture() {
 		drm_close_driver(fd);
 	}
-- 
2.47.3

Re: [PATCH 4/4] panthor: Add simple compute-shader test

[Kamil Konieczny]

Hi Daniel,
On 2026-10-15 at 14:05:49 +0100, Daniel Stone wrote:
> This test launches a compute shader which performs RMW on a memory
> array, to check that we can actually launch jobs and see the results.
> 
> Signed-off-by: Daniel Stone <daniels@collabora.com>
> ---
>  lib/igt_panthor.h             | 204 +++++++++++++++++++++++++++++
>  tests/panthor/panthor_group.c | 240 ++++++++++++++++++++++++++++++++++
>  2 files changed, 444 insertions(+)
> 
> diff --git a/lib/igt_panthor.h b/lib/igt_panthor.h
> index dfeb0ac60..35882fad7 100644
> --- a/lib/igt_panthor.h
> +++ b/lib/igt_panthor.h
> @@ -363,4 +363,208 @@ cs_error_barrier(void)
>  	return instr.raw;
>  }
>  
> +struct gpu_src_reg {
> +	uint8_t reg;
> +	bool last;
> +	uint8_t rep;
> +};
> +
> +struct gpu_dst_reg {
> +	uint8_t reg;
> +};
> +
> +enum gpu_flow_ctrl {
> +	FLOW_NONE = 0,
> +	FLOW_WAIT0 = 0x1,
> +	FLOW_WAIT1 = 0x2,
> +	FLOW_WAIT01 = 0x3,
> +	FLOW_WAIT2 = 0x4,
> +	FLOW_WAIT02 = 0x5,
> +	FLOW_WAIT12 = 0x6,
> +	FLOW_WAIT012 = 0x7,
> +	FLOW_WAIT0126 = 0x8,
> +	FLOW_WAIT = 0x9,
> +	FLOW_RECONVERGE = 0xa,
> +	FLOW_WAIT_RESOURCE = 0xc,
> +	FLOW_DISCARD = 0xd,
> +	FLOW_END = 0xf,
> +};
> +
> +enum gpu_reg_width {
> +	REG_WIDTH_LO16 = 0x1,
> +	REG_WIDTH_HI16 = 0x2,
> +	REG_WIDTH_32 = 0x3,
> +};
> +
> +static inline struct gpu_dst_reg
> +gpu_dst_reg(uint8_t reg)
> +{
> +	struct gpu_dst_reg ret = {
> +		.reg = reg
> +	};
> +	return ret;
> +}
> +
> +static inline struct gpu_src_reg
> +gpu_src_zero(void)
> +{
> +	struct gpu_src_reg ret = {
> +		.reg = 255, /* architecturally invalid */
> +		.last = false,
> +		.rep = (0x3 << 6),
> +	};
> +	return ret;
> +}
> +
> +static inline struct gpu_src_reg
> +gpu_src_reg(uint8_t reg)
> +{
> +	struct gpu_src_reg ret = {
> +		.reg = reg,
> +		.last = false,
> +		.rep = reg,
> +	};
> +	return ret;
> +}
> +
> +static inline struct gpu_src_reg
> +gpu_src_reg_last(uint8_t reg)
> +{
> +	struct gpu_src_reg ret = {
> +		.reg = reg,
> +		.last = true,
> +		.rep = (reg | (1 << 6)),
> +	};
> +	return ret;
> +}
> +
> +#define ins(x, v, s) x |= ((uint64_t) v << s)

imho this should be:
#define ins(x, v, s) x |= (((uint64_t)(v)) << (s))

See warning from checkpatch.pl script. Warn about x could be ignored.

> +
> +static inline uint64_t
> +gpu_mov_imm_i32(struct gpu_dst_reg dst, int32_t imm, enum gpu_flow_ctrl flow)
> +{
> +	uint64_t ret = 0;

Add newline.

> +	ins(ret, gpu_src_zero().rep, 0);
> +	ins(ret, imm, 8);
> +	ins(ret, dst.reg, 40);
> +	ins(ret, REG_WIDTH_32, 46);
> +	ins(ret, 0x0, 48); /* opcode */
> +	ins(ret, 0x11, 52); /* class */
> +	ins(ret, 0x0, 57); /* FAU */
> +	ins(ret, flow, 59);

Add newline.

> +	return ret;
> +}
> +
> +static inline uint64_t
> +gpu_lshift_or_i32(struct gpu_dst_reg dst, struct gpu_src_reg src,
> +		  struct gpu_src_reg shift, struct gpu_src_reg or,
> +		  enum gpu_flow_ctrl flow)
> +{
> +	uint64_t ret = 0;

Add newline.

> +	ins(ret, src.rep, 0);
> +	ins(ret, shift.rep, 8);
> +	ins(ret, or.rep, 16);
> +	ins(ret, 0x1, 24); /* logicop = or */
> +	ins(ret, 0x4, 26); /* lanes = b0 */
> +	ins(ret, 0x0, 30); /* dst.not = false */
> +	ins(ret, 0x1, 32); /* shift = left */
> +	ins(ret, 0x0, 35); /* src2.not = false */
> +	ins(ret, 0x0, 36); /* widen = false */
> +	ins(ret, dst.reg, 40);
> +	ins(ret, REG_WIDTH_32, 46);
> +	ins(ret, 0x4, 48); /* opcode */
> +	ins(ret, 0xb, 52); /* class */
> +	ins(ret, 0x0, 57); /* FAU */
> +	ins(ret, flow, 59);

Add newline.

> +	return ret;
> +}
> +
> +static inline uint64_t
> +gpu_iadd_u32(struct gpu_dst_reg dst, struct gpu_src_reg left,
> +	     struct gpu_src_reg right, enum gpu_flow_ctrl flow)
> +{
> +	uint64_t ret = 0;

Add newline.

> +	ins(ret, left.rep, 0);
> +	ins(ret, right.rep, 8);
> +	ins(ret, 0x0, 16); /* opcode */
> +	ins(ret, 0x0, 26); /* widen1 */
> +	ins(ret, 0x0, 30); /* saturate */
> +	ins(ret, 0x0, 36); /* widen0 */
> +	ins(ret, dst.reg, 40);
> +	ins(ret, REG_WIDTH_32, 46);
> +	ins(ret, 0x0, 48); /* data type */
> +	ins(ret, 0xa, 52); /* class */
> +	ins(ret, 0x0, 57); /* FAU */
> +	ins(ret, flow, 59);

Add newline.

> +	return ret;
> +}
> +
> +static inline uint64_t
> +gpu_imul_u32(struct gpu_dst_reg dst, struct gpu_src_reg left,
> +	     struct gpu_src_reg right, enum gpu_flow_ctrl flow)
> +{
> +	uint64_t ret = 0;

Add newline.

> +	ins(ret, left.rep, 0);
> +	ins(ret, right.rep, 8);
> +	ins(ret, 0xa, 16); /* opcode */
> +	ins(ret, 0x0, 26); /* widen1 */
> +	ins(ret, 0x0, 30); /* saturate */
> +	ins(ret, 0x0, 36); /* widen0 */
> +	ins(ret, dst.reg, 40);
> +	ins(ret, REG_WIDTH_32, 46);
> +	ins(ret, 0x0, 48); /* data type */
> +	ins(ret, 0xa, 52); /* class */
> +	ins(ret, 0x0, 57); /* FAU */
> +	ins(ret, flow, 59);

Add newline.

> +	return ret;
> +}
> +
> +static inline uint64_t
> +gpu_load_i32(struct gpu_dst_reg dst, struct gpu_src_reg addr_lo,
> +	     struct gpu_src_reg addr_hi, uint16_t offset,
> +	     uint8_t sb_slot, enum gpu_flow_ctrl flow)
> +{
> +	uint64_t ret = 0;

Add newline.

> +	igt_assert_eq(addr_hi.reg, addr_lo.reg + 1);
> +	ins(ret, addr_lo.rep, 0);
> +	ins(ret, offset, 8);
> +	ins(ret, 0x0, 24); /* access mode */
> +	ins(ret, REG_WIDTH_32, 27);
> +	ins(ret, sb_slot, 30);
> +	ins(ret, 0x1, 33); /* staging register count */
> +	ins(ret, 0x0, 36); /* load 32 */
> +	ins(ret, 0x1, 39); /* no sign extension */
> +	ins(ret, dst.reg, 40); /* base staging register */
> +	ins(ret, 0x2, 46); /* write staging registers */
> +	ins(ret, 0x0, 48); /* opcode */
> +	ins(ret, 0x6, 52); /* class */
> +	ins(ret, 0x0, 57); /* FAU */
> +	ins(ret, flow, 59);

Add newline.

> +	return ret;
> +}
> +
> +static inline uint64_t
> +gpu_store_i32(struct gpu_src_reg addr_lo, struct gpu_src_reg addr_hi,
> +	      struct gpu_src_reg val, uint16_t offset,
> +	      uint8_t sb_slot, enum gpu_flow_ctrl flow)
> +{
> +	uint64_t ret = 0;

Add newline.

> +	igt_assert_eq(addr_hi.reg, addr_lo.reg + 1);
> +	ins(ret, addr_lo.rep, 0);
> +	ins(ret, offset, 8);
> +	ins(ret, 0x0, 24); /* access mode */
> +	ins(ret, REG_WIDTH_32, 27);
> +	ins(ret, sb_slot, 30);
> +	ins(ret, 0x1, 33); /* staging register count */
> +	ins(ret, val.reg, 40); /* base staging register */
> +	ins(ret, 0x1, 46); /* read staging registers */
> +	ins(ret, 0x1, 48); /* opcode */
> +	ins(ret, 0x6, 52); /* class */
> +	ins(ret, 0x0, 57); /* FAU */
> +	ins(ret, flow, 59);

Add newline.

> +	return ret;
> +}
> +
> +#undef ins
> +
>  #endif /* IGT_PANTHOR_H */
> diff --git a/tests/panthor/panthor_group.c b/tests/panthor/panthor_group.c
> index c79f8d629..feca42482 100644
> --- a/tests/panthor/panthor_group.c
> +++ b/tests/panthor/panthor_group.c
> @@ -10,6 +10,133 @@
>  #include "igt_syncobj.h"
>  #include "panthor_drm.h"
>  
> +/*
> + * Record a simple compute program:
> + * {
> + *     extern uint32_t shared_buf[num_invocations];
> + *     extern const uint32_t my_invocation_id;
> + *     shared_buf[my_invocation_id] *= my_invocation_id;
> + * }
> + */
> +static void
> +record_compute_prog(uint8_t *csf_cmdbuf, unsigned int *csf_cmdbuf_size,
> +		    uint8_t *gpu_cmdbuf, unsigned int *gpu_cmdbuf_size,
> +		    uint32_t srt_gpuva, uint32_t fau_gpuva, uint32_t spd_gpuva,
> +		    uint32_t tsd_gpuva, uint32_t data_gpuva)
> +{
> +	const uint8_t reg_id_global_x = 60; /* preloaded */
> +	const uint8_t reg_addr_lo = 0;
> +	const uint8_t reg_addr_hi = 1;
> +	const uint8_t reg_val = 2;
> +	const uint8_t reg_offset = 3;
> +	const uint8_t reg_shift = 4;
> +	const uint8_t compute_sr_base = 0;
> +	const uint8_t zero_reg = 40;
> +
> +	uint64_t gpu_instrs[] = {
> +		/* MOV_IMM.i32 rADDRLO, 0x10000 */
> +		gpu_mov_imm_i32(gpu_dst_reg(reg_addr_lo), data_gpuva, FLOW_NONE),
> +		/* MOV_IMM.i32 rADDRHI, 0x0 */
> +		gpu_mov_imm_i32(gpu_dst_reg(reg_addr_hi), 0x0, FLOW_NONE),
> +		/* MOV_IMM.i32 rSHIFT, 0x2 */
> +		gpu_mov_imm_i32(gpu_dst_reg(reg_shift), 0x2, FLOW_NONE),
> +		/* LSHIFT_OR.i32 rOFFSET, rGLOBALID, rSHIFT^, 0x0 */
> +		gpu_lshift_or_i32(gpu_dst_reg(reg_offset),
> +				  gpu_src_reg(reg_id_global_x),
> +				  gpu_src_reg_last(reg_shift),
> +				  gpu_src_zero(),
> +				  FLOW_NONE),
> +		/* IADD.u32 rADDRLO, rADDRLO^, rOFFSET^ */
> +		gpu_iadd_u32(gpu_dst_reg(reg_addr_lo),
> +			     gpu_src_reg_last(reg_addr_lo),
> +			     gpu_src_reg_last(reg_offset),
> +			     FLOW_NONE),
> +		/* LOAD.i32.slot0.wait0 @rVAL, [rADDRLO:rADDRHI], offset:0 */
> +		gpu_load_i32(gpu_dst_reg(reg_val),
> +			     gpu_src_reg(reg_addr_lo),
> +			     gpu_src_reg(reg_addr_hi),
> +			     0x0,
> +			     0, FLOW_WAIT0),
> +		/* IMUL_u32 rVAL, rVAL^, rGLOBALID^ */
> +		gpu_imul_u32(gpu_dst_reg(reg_val),
> +			     gpu_src_reg_last(reg_val),
> +			     gpu_src_reg_last(reg_id_global_x),
> +			     FLOW_NONE),
> +		/* STORE.i32.slot1.end @rVAL, [rADDRLO^:rADDRHI^], offset:0 */
> +		gpu_store_i32(gpu_src_reg_last(reg_addr_lo),
> +			      gpu_src_reg_last(reg_addr_hi),
> +			      gpu_src_reg(reg_val),
> +			      0x0,
> +			      1, FLOW_END),
> +	};
> +
> +	uint64_t csf_instrs[] = {
> +		/* REQ_RESOURCE.compute */
> +		cs_req_resource(true, false, false, false),
> +		/* MOV32 r40, 0 (always force cache flush/invalidate) */
> +		cs_mov32(zero_reg, 0),
> +		/* FLUSH_CACHE.clean_invalidate_l2.clean_invalidate_lsc.invalidate_other.sb0 */
> +		cs_flush(CS_FLUSH_MODE_CLEAN_AND_INVALIDATE,
> +			 CS_FLUSH_MODE_CLEAN_AND_INVALIDATE,
> +			 CS_FLUSH_MODE_INVALIDATE,
> +			 0, zero_reg, 1),
> +		/* WAIT (flush) */
> +		cs_wait(0xff, false),
> +		/* ShaderResourceTable 0-3 pointers */
> +		cs_mov48(compute_sr_base +  0, srt_gpuva),
> +		cs_mov48(compute_sr_base +  2, 0x222222222222),
> +		cs_mov48(compute_sr_base +  4, 0x333333333333),
> +		cs_mov48(compute_sr_base +  6, 0x444444444444),
> +		/* FastAccessUniform table 0-3 pointers */
> +		cs_mov48(compute_sr_base +  8, fau_gpuva),
> +		cs_mov48(compute_sr_base + 10, 0x666666666666),
> +		cs_mov48(compute_sr_base + 12, 0x777777777777),
> +		cs_mov48(compute_sr_base + 14, 0x888888888888),
> +		/* ShaderProgramDescriptor 0-3 pointers */
> +		cs_mov48(compute_sr_base + 16, spd_gpuva),
> +		cs_mov48(compute_sr_base + 18, 0x999999999999),
> +		cs_mov48(compute_sr_base + 20, 0xaaaaaaaaaaaa),
> +		cs_mov48(compute_sr_base + 22, 0xbbbbbbbbbbbb),
> +		/* ThreadStorageDescriptor 0-3 pointers */
> +		cs_mov48(compute_sr_base + 24, tsd_gpuva),
> +		cs_mov48(compute_sr_base + 26, 0xdddddddddddd),
> +		cs_mov48(compute_sr_base + 28, 0xeeeeeeeeeeee),
> +		cs_mov48(compute_sr_base + 30, 0xffffffffffff),
> +		/* global ID offset = 0 */
> +		cs_mov32(compute_sr_base + 32, 0),
> +		/* workgroup size 64x1x1, allow merge */
> +		cs_mov32(compute_sr_base + 33, (63 << 0) | (0 << 10) | (0 << 20) | (1 << 31)),
> +		/* job offset (x/y/z) = { 0, 0, 0 } */
> +		cs_mov32(compute_sr_base + 34, 0),
> +		cs_mov32(compute_sr_base + 35, 0),
> +		cs_mov32(compute_sr_base + 36, 0),
> +		/* job size (x/y/z) = { 64, 1, 1 } */
> +		cs_mov32(compute_sr_base + 37, 64),
> +		cs_mov32(compute_sr_base + 38, 1),
> +		cs_mov32(compute_sr_base + 39, 1),
> +		/* SET_SB_ENTRY.async_ep1.other1 */
> +		cs_set_sb_entry(1, 1),
> +		/* RUN_COMPUTE.increment64.x_axis.no_increment.srt0.spd0.tsd0.fau0 */
> +		cs_run_compute(64, CS_TASK_AXIS_X, 0, 0, 0, 0, 0),
> +		/* WAIT (RUN_COMPUTE) */
> +		cs_wait(0xff, false),
> +		/* FLUSH_CACHE.clean_invalidate_l2.clean_invalidate_lsc.invalidate_other.sb0 */
> +		cs_flush(CS_FLUSH_MODE_CLEAN_AND_INVALIDATE,
> +			 CS_FLUSH_MODE_CLEAN_AND_INVALIDATE,
> +			 CS_FLUSH_MODE_INVALIDATE,
> +			 0, zero_reg, 1),
> +		/* WAIT (flush) */
> +		cs_wait(0xff, false),
> +		/* ERROR_BARRIER */
> +		cs_error_barrier(),
> +	};
> +
> +	memcpy(gpu_cmdbuf, gpu_instrs, sizeof(gpu_instrs));
> +	*gpu_cmdbuf_size = sizeof(gpu_instrs);
> +	memcpy(csf_cmdbuf, csf_instrs, sizeof(csf_instrs));
> +	*csf_cmdbuf_size = sizeof(csf_instrs);
> +}
> +
>  static size_t
>  issue_store_multiple(uint8_t *cs, uint64_t kernel_va, uint32_t constant)
>  {
> @@ -118,6 +245,119 @@ int igt_main() {
>  		igt_panthor_free_bo(fd, &result_bo);
>  	}
>  
> +	igt_describe("Submit a simple compute job with memory access.");
> +	igt_subtest("simple_compute") {
> +		uint32_t vm_id;
> +		uint32_t group_handle;
> +		struct panthor_bo csf_cmdbuf_bo = {};
> +		struct panthor_bo gpu_cmdbuf_bo = {};
> +		struct panthor_bo resource_table_bo = {};
> +		struct panthor_bo data_bo = {};
> +		uint64_t csf_cmdbuf_gpuva;
> +		uint32_t csf_cmdbuf_size;
> +		uint64_t gpu_cmdbuf_gpuva;
> +		uint32_t gpu_cmdbuf_size;
> +		uint32_t resource_table_gpuva;
> +		uint32_t data_gpuva;
> +		uint32_t syncobj_handle;
> +		uint32_t srt_offset;
> +		uint32_t fau_offset;
> +		uint32_t spd_offset;
> +		uint32_t *spd_ptr;
> +		uint32_t spd_gpuva;
> +		uint32_t tsd_offset;
> +		uint32_t *tsd_ptr;
> +		int arch = igt_panthor_get_arch(fd);
> +
> +		igt_require(arch < 15);
> +
> +		igt_panthor_vm_create(fd, &vm_id, 0);
> +
> +		/* CSF ISA command stream */
> +		csf_cmdbuf_gpuva = 0x1000000;
> +		igt_panthor_bo_create_mapped(fd, &csf_cmdbuf_bo, 4096, 0, 0);
> +		igt_panthor_vm_bind(fd, vm_id, csf_cmdbuf_bo.handle, csf_cmdbuf_gpuva,
> +				    csf_cmdbuf_bo.size, DRM_PANTHOR_VM_BIND_OP_TYPE_MAP, 0);
> +		memset(csf_cmdbuf_bo.map, 0, csf_cmdbuf_bo.size);
> +
> +		/* GPU ISA command stream */
> +		gpu_cmdbuf_gpuva = 0x2000000;
> +		igt_panthor_bo_create_mapped(fd, &gpu_cmdbuf_bo, 4096, 0, 0);
> +		igt_panthor_vm_bind(fd, vm_id, gpu_cmdbuf_bo.handle, gpu_cmdbuf_gpuva,
> +				    gpu_cmdbuf_bo.size, DRM_PANTHOR_VM_BIND_OP_TYPE_MAP, 0);
> +		memset(gpu_cmdbuf_bo.map, 0, gpu_cmdbuf_bo.size);
> +
> +		/* resource tables passed to RUN_COMPUTE */
> +		resource_table_gpuva = 0x3000000;
> +		igt_panthor_bo_create_mapped(fd, &resource_table_bo, 4096, 0, 0);
> +		igt_panthor_vm_bind(fd, vm_id, resource_table_bo.handle, resource_table_gpuva,
> +				    resource_table_bo.size, DRM_PANTHOR_VM_BIND_OP_TYPE_MAP, 0);
> +		memset(resource_table_bo.map, 0, resource_table_bo.size);
> +		/* ShaderResourceTable descriptor (zeroed) */
> +		srt_offset = 0;
> +		/* FastAccessUniform descriptor (zeroed) */
> +		fau_offset = 128;
> +		/* ShaderProgramDescriptor */
> +		spd_offset = 256;
> +		spd_gpuva = resource_table_gpuva + spd_offset;
> +		if (arch == 14)
> +			spd_gpuva |= 2; /* register allocation propagated into ptr */
> +		spd_ptr = (uint32_t *) ((unsigned char *) resource_table_bo.map + spd_offset);

No need for space after a cast:

		spd_ptr = (uint32_t *)((unsigned char *)resource_table_bo.map + spd_offset);

> +		spd_ptr[0] = 0x8 | (0x1 << 4); /* compute shader */
> +		spd_ptr[0] |= (0 << 16) | (0 << 17) | (0 << 19); /* no NaN suppress, FTZ, inf suppress */

You can ignore checkpatch warn 'line too long'.

> +		spd_ptr[0] |= (0 << 29) | (2 << 30); /* no JUMP_EX, max 32 registers */
> +		spd_ptr[1] |= (1 << 12); /* preload GlobalInvocationID.x into r60 */
> +		spd_ptr[2] = gpu_cmdbuf_gpuva; /* initial PC */
> +		/* ThreadStorageDescriptor */
> +		tsd_offset = 512;
> +		tsd_ptr = (uint32_t *) ((unsigned char *) resource_table_bo.map + tsd_offset);
> +		tsd_ptr[1] = 0x1f; /* max wls_instances */
> +
> +		/* CPU<->GPU shared data */
> +		data_gpuva = 0x4000000;
> +		igt_panthor_bo_create_mapped(fd, &data_bo, 16384, 0, 0);
> +		igt_panthor_vm_bind(fd, vm_id, data_bo.handle, data_gpuva,
> +				    data_bo.size, DRM_PANTHOR_VM_BIND_OP_TYPE_MAP, 0);
> +		for (unsigned int i = 0; i < 64 * 64; i++) {
> +			uint32_t *x = data_bo.map + (i * sizeof(uint32_t));
> +			*x = i;
> +		}
> +
> +		record_compute_prog(csf_cmdbuf_bo.map, &csf_cmdbuf_size,
> +				    gpu_cmdbuf_bo.map, &gpu_cmdbuf_size,
> +				    resource_table_gpuva + srt_offset,
> +				    resource_table_gpuva + fau_offset,
> +				    spd_gpuva,
> +				    resource_table_gpuva + tsd_offset,
> +				    data_gpuva);
> +
> +		group_handle = igt_panthor_group_create_simple(fd, vm_id, 0);
> +		igt_assert_neq(group_handle, 0);
> +
> +		syncobj_handle = syncobj_create(fd, 0);
> +
> +		igt_panthor_group_submit_simple(fd, group_handle, 0,
> +						csf_cmdbuf_gpuva, csf_cmdbuf_size,
> +						syncobj_handle, 0);
> +		igt_assert(syncobj_wait(fd, &syncobj_handle, 1, INT64_MAX, 0, NULL));
> +
> +		for (unsigned int i = 0; i < 64 * 64; i++) {
> +			uint32_t *x = data_bo.map + (i * sizeof(uint32_t));

Add newline.

With above fixed this is
Acked-by: Kamil Konieczny <kamil.konieczny@linux.intel.com>

Regards,
Kamil

> +			igt_assert_eq(*x, i * i);
> +		}
> +
> +		syncobj_destroy(fd, syncobj_handle);
> +
> +		igt_panthor_group_destroy(fd, group_handle, 0);
> +
> +		igt_panthor_vm_destroy(fd, vm_id, 0);
> +
> +		igt_panthor_free_bo(fd, &data_bo);
> +		igt_panthor_free_bo(fd, &resource_table_bo);
> +		igt_panthor_free_bo(fd, &gpu_cmdbuf_bo);
> +		igt_panthor_free_bo(fd, &csf_cmdbuf_bo);
> +	}
> +
>  	igt_fixture() {
>  		drm_close_driver(fd);
>  	}
> -- 
> 2.47.3
> 

Re: [PATCH 4/4] panthor: Add simple compute-shader test

[Liviu Dudau]

On Thu, Oct 15, 2026 at 02:05:49PM +0100, Daniel Stone wrote:
> This test launches a compute shader which performs RMW on a memory
> array, to check that we can actually launch jobs and see the results.
> 
> Signed-off-by: Daniel Stone <daniels@collabora.com>
> ---
>  lib/igt_panthor.h             | 204 +++++++++++++++++++++++++++++
>  tests/panthor/panthor_group.c | 240 ++++++++++++++++++++++++++++++++++
>  2 files changed, 444 insertions(+)
> 
> diff --git a/lib/igt_panthor.h b/lib/igt_panthor.h
> index dfeb0ac60..35882fad7 100644
> --- a/lib/igt_panthor.h
> +++ b/lib/igt_panthor.h
> @@ -363,4 +363,208 @@ cs_error_barrier(void)
>  	return instr.raw;
>  }
>  
> +struct gpu_src_reg {
> +	uint8_t reg;
> +	bool last;
> +	uint8_t rep;
> +};
> +
> +struct gpu_dst_reg {
> +	uint8_t reg;
> +};
> +
> +enum gpu_flow_ctrl {
> +	FLOW_NONE = 0,
> +	FLOW_WAIT0 = 0x1,
> +	FLOW_WAIT1 = 0x2,
> +	FLOW_WAIT01 = 0x3,
> +	FLOW_WAIT2 = 0x4,
> +	FLOW_WAIT02 = 0x5,
> +	FLOW_WAIT12 = 0x6,
> +	FLOW_WAIT012 = 0x7,
> +	FLOW_WAIT0126 = 0x8,
> +	FLOW_WAIT = 0x9,
> +	FLOW_RECONVERGE = 0xa,
> +	FLOW_WAIT_RESOURCE = 0xc,
> +	FLOW_DISCARD = 0xd,
> +	FLOW_END = 0xf,
> +};
> +
> +enum gpu_reg_width {
> +	REG_WIDTH_LO16 = 0x1,
> +	REG_WIDTH_HI16 = 0x2,
> +	REG_WIDTH_32 = 0x3,
> +};
> +
> +static inline struct gpu_dst_reg
> +gpu_dst_reg(uint8_t reg)
> +{
> +	struct gpu_dst_reg ret = {
> +		.reg = reg
> +	};
> +	return ret;
> +}
> +
> +static inline struct gpu_src_reg
> +gpu_src_zero(void)
> +{
> +	struct gpu_src_reg ret = {
> +		.reg = 255, /* architecturally invalid */
> +		.last = false,
> +		.rep = (0x3 << 6),
> +	};
> +	return ret;
> +}
> +
> +static inline struct gpu_src_reg
> +gpu_src_reg(uint8_t reg)
> +{
> +	struct gpu_src_reg ret = {
> +		.reg = reg,
> +		.last = false,
> +		.rep = reg,
> +	};
> +	return ret;
> +}
> +
> +static inline struct gpu_src_reg
> +gpu_src_reg_last(uint8_t reg)
> +{
> +	struct gpu_src_reg ret = {
> +		.reg = reg,
> +		.last = true,
> +		.rep = (reg | (1 << 6)),
> +	};
> +	return ret;
> +}
> +
> +#define ins(x, v, s) x |= ((uint64_t) v << s)

Minor nit to add to Kamil's comments: can you also indent the content of the define a bit more?

Otherwise,

Reviewed-by: Liviu Dudau <liviu.dudau@arm.com>

Best regards,
Liviu


> +
> +static inline uint64_t
> +gpu_mov_imm_i32(struct gpu_dst_reg dst, int32_t imm, enum gpu_flow_ctrl flow)
> +{
> +	uint64_t ret = 0;
> +	ins(ret, gpu_src_zero().rep, 0);
> +	ins(ret, imm, 8);
> +	ins(ret, dst.reg, 40);
> +	ins(ret, REG_WIDTH_32, 46);
> +	ins(ret, 0x0, 48); /* opcode */
> +	ins(ret, 0x11, 52); /* class */
> +	ins(ret, 0x0, 57); /* FAU */
> +	ins(ret, flow, 59);
> +	return ret;
> +}
> +
> +static inline uint64_t
> +gpu_lshift_or_i32(struct gpu_dst_reg dst, struct gpu_src_reg src,
> +		  struct gpu_src_reg shift, struct gpu_src_reg or,
> +		  enum gpu_flow_ctrl flow)
> +{
> +	uint64_t ret = 0;
> +	ins(ret, src.rep, 0);
> +	ins(ret, shift.rep, 8);
> +	ins(ret, or.rep, 16);
> +	ins(ret, 0x1, 24); /* logicop = or */
> +	ins(ret, 0x4, 26); /* lanes = b0 */
> +	ins(ret, 0x0, 30); /* dst.not = false */
> +	ins(ret, 0x1, 32); /* shift = left */
> +	ins(ret, 0x0, 35); /* src2.not = false */
> +	ins(ret, 0x0, 36); /* widen = false */
> +	ins(ret, dst.reg, 40);
> +	ins(ret, REG_WIDTH_32, 46);
> +	ins(ret, 0x4, 48); /* opcode */
> +	ins(ret, 0xb, 52); /* class */
> +	ins(ret, 0x0, 57); /* FAU */
> +	ins(ret, flow, 59);
> +	return ret;
> +}
> +
> +static inline uint64_t
> +gpu_iadd_u32(struct gpu_dst_reg dst, struct gpu_src_reg left,
> +	     struct gpu_src_reg right, enum gpu_flow_ctrl flow)
> +{
> +	uint64_t ret = 0;
> +	ins(ret, left.rep, 0);
> +	ins(ret, right.rep, 8);
> +	ins(ret, 0x0, 16); /* opcode */
> +	ins(ret, 0x0, 26); /* widen1 */
> +	ins(ret, 0x0, 30); /* saturate */
> +	ins(ret, 0x0, 36); /* widen0 */
> +	ins(ret, dst.reg, 40);
> +	ins(ret, REG_WIDTH_32, 46);
> +	ins(ret, 0x0, 48); /* data type */
> +	ins(ret, 0xa, 52); /* class */
> +	ins(ret, 0x0, 57); /* FAU */
> +	ins(ret, flow, 59);
> +	return ret;
> +}
> +
> +static inline uint64_t
> +gpu_imul_u32(struct gpu_dst_reg dst, struct gpu_src_reg left,
> +	     struct gpu_src_reg right, enum gpu_flow_ctrl flow)
> +{
> +	uint64_t ret = 0;
> +	ins(ret, left.rep, 0);
> +	ins(ret, right.rep, 8);
> +	ins(ret, 0xa, 16); /* opcode */
> +	ins(ret, 0x0, 26); /* widen1 */
> +	ins(ret, 0x0, 30); /* saturate */
> +	ins(ret, 0x0, 36); /* widen0 */
> +	ins(ret, dst.reg, 40);
> +	ins(ret, REG_WIDTH_32, 46);
> +	ins(ret, 0x0, 48); /* data type */
> +	ins(ret, 0xa, 52); /* class */
> +	ins(ret, 0x0, 57); /* FAU */
> +	ins(ret, flow, 59);
> +	return ret;
> +}
> +
> +static inline uint64_t
> +gpu_load_i32(struct gpu_dst_reg dst, struct gpu_src_reg addr_lo,
> +	     struct gpu_src_reg addr_hi, uint16_t offset,
> +	     uint8_t sb_slot, enum gpu_flow_ctrl flow)
> +{
> +	uint64_t ret = 0;
> +	igt_assert_eq(addr_hi.reg, addr_lo.reg + 1);
> +	ins(ret, addr_lo.rep, 0);
> +	ins(ret, offset, 8);
> +	ins(ret, 0x0, 24); /* access mode */
> +	ins(ret, REG_WIDTH_32, 27);
> +	ins(ret, sb_slot, 30);
> +	ins(ret, 0x1, 33); /* staging register count */
> +	ins(ret, 0x0, 36); /* load 32 */
> +	ins(ret, 0x1, 39); /* no sign extension */
> +	ins(ret, dst.reg, 40); /* base staging register */
> +	ins(ret, 0x2, 46); /* write staging registers */
> +	ins(ret, 0x0, 48); /* opcode */
> +	ins(ret, 0x6, 52); /* class */
> +	ins(ret, 0x0, 57); /* FAU */
> +	ins(ret, flow, 59);
> +	return ret;
> +}
> +
> +static inline uint64_t
> +gpu_store_i32(struct gpu_src_reg addr_lo, struct gpu_src_reg addr_hi,
> +	      struct gpu_src_reg val, uint16_t offset,
> +	      uint8_t sb_slot, enum gpu_flow_ctrl flow)
> +{
> +	uint64_t ret = 0;
> +	igt_assert_eq(addr_hi.reg, addr_lo.reg + 1);
> +	ins(ret, addr_lo.rep, 0);
> +	ins(ret, offset, 8);
> +	ins(ret, 0x0, 24); /* access mode */
> +	ins(ret, REG_WIDTH_32, 27);
> +	ins(ret, sb_slot, 30);
> +	ins(ret, 0x1, 33); /* staging register count */
> +	ins(ret, val.reg, 40); /* base staging register */
> +	ins(ret, 0x1, 46); /* read staging registers */
> +	ins(ret, 0x1, 48); /* opcode */
> +	ins(ret, 0x6, 52); /* class */
> +	ins(ret, 0x0, 57); /* FAU */
> +	ins(ret, flow, 59);
> +	return ret;
> +}
> +
> +#undef ins
> +
>  #endif /* IGT_PANTHOR_H */
> diff --git a/tests/panthor/panthor_group.c b/tests/panthor/panthor_group.c
> index c79f8d629..feca42482 100644
> --- a/tests/panthor/panthor_group.c
> +++ b/tests/panthor/panthor_group.c
> @@ -10,6 +10,133 @@
>  #include "igt_syncobj.h"
>  #include "panthor_drm.h"
>  
> +/*
> + * Record a simple compute program:
> + * {
> + *     extern uint32_t shared_buf[num_invocations];
> + *     extern const uint32_t my_invocation_id;
> + *     shared_buf[my_invocation_id] *= my_invocation_id;
> + * }
> + */
> +static void
> +record_compute_prog(uint8_t *csf_cmdbuf, unsigned int *csf_cmdbuf_size,
> +		    uint8_t *gpu_cmdbuf, unsigned int *gpu_cmdbuf_size,
> +		    uint32_t srt_gpuva, uint32_t fau_gpuva, uint32_t spd_gpuva,
> +		    uint32_t tsd_gpuva, uint32_t data_gpuva)
> +{
> +	const uint8_t reg_id_global_x = 60; /* preloaded */
> +	const uint8_t reg_addr_lo = 0;
> +	const uint8_t reg_addr_hi = 1;
> +	const uint8_t reg_val = 2;
> +	const uint8_t reg_offset = 3;
> +	const uint8_t reg_shift = 4;
> +	const uint8_t compute_sr_base = 0;
> +	const uint8_t zero_reg = 40;
> +
> +	uint64_t gpu_instrs[] = {
> +		/* MOV_IMM.i32 rADDRLO, 0x10000 */
> +		gpu_mov_imm_i32(gpu_dst_reg(reg_addr_lo), data_gpuva, FLOW_NONE),
> +		/* MOV_IMM.i32 rADDRHI, 0x0 */
> +		gpu_mov_imm_i32(gpu_dst_reg(reg_addr_hi), 0x0, FLOW_NONE),
> +		/* MOV_IMM.i32 rSHIFT, 0x2 */
> +		gpu_mov_imm_i32(gpu_dst_reg(reg_shift), 0x2, FLOW_NONE),
> +		/* LSHIFT_OR.i32 rOFFSET, rGLOBALID, rSHIFT^, 0x0 */
> +		gpu_lshift_or_i32(gpu_dst_reg(reg_offset),
> +				  gpu_src_reg(reg_id_global_x),
> +				  gpu_src_reg_last(reg_shift),
> +				  gpu_src_zero(),
> +				  FLOW_NONE),
> +		/* IADD.u32 rADDRLO, rADDRLO^, rOFFSET^ */
> +		gpu_iadd_u32(gpu_dst_reg(reg_addr_lo),
> +			     gpu_src_reg_last(reg_addr_lo),
> +			     gpu_src_reg_last(reg_offset),
> +			     FLOW_NONE),
> +		/* LOAD.i32.slot0.wait0 @rVAL, [rADDRLO:rADDRHI], offset:0 */
> +		gpu_load_i32(gpu_dst_reg(reg_val),
> +			     gpu_src_reg(reg_addr_lo),
> +			     gpu_src_reg(reg_addr_hi),
> +			     0x0,
> +			     0, FLOW_WAIT0),
> +		/* IMUL_u32 rVAL, rVAL^, rGLOBALID^ */
> +		gpu_imul_u32(gpu_dst_reg(reg_val),
> +			     gpu_src_reg_last(reg_val),
> +			     gpu_src_reg_last(reg_id_global_x),
> +			     FLOW_NONE),
> +		/* STORE.i32.slot1.end @rVAL, [rADDRLO^:rADDRHI^], offset:0 */
> +		gpu_store_i32(gpu_src_reg_last(reg_addr_lo),
> +			      gpu_src_reg_last(reg_addr_hi),
> +			      gpu_src_reg(reg_val),
> +			      0x0,
> +			      1, FLOW_END),
> +	};
> +
> +	uint64_t csf_instrs[] = {
> +		/* REQ_RESOURCE.compute */
> +		cs_req_resource(true, false, false, false),
> +		/* MOV32 r40, 0 (always force cache flush/invalidate) */
> +		cs_mov32(zero_reg, 0),
> +		/* FLUSH_CACHE.clean_invalidate_l2.clean_invalidate_lsc.invalidate_other.sb0 */
> +		cs_flush(CS_FLUSH_MODE_CLEAN_AND_INVALIDATE,
> +			 CS_FLUSH_MODE_CLEAN_AND_INVALIDATE,
> +			 CS_FLUSH_MODE_INVALIDATE,
> +			 0, zero_reg, 1),
> +		/* WAIT (flush) */
> +		cs_wait(0xff, false),
> +		/* ShaderResourceTable 0-3 pointers */
> +		cs_mov48(compute_sr_base +  0, srt_gpuva),
> +		cs_mov48(compute_sr_base +  2, 0x222222222222),
> +		cs_mov48(compute_sr_base +  4, 0x333333333333),
> +		cs_mov48(compute_sr_base +  6, 0x444444444444),
> +		/* FastAccessUniform table 0-3 pointers */
> +		cs_mov48(compute_sr_base +  8, fau_gpuva),
> +		cs_mov48(compute_sr_base + 10, 0x666666666666),
> +		cs_mov48(compute_sr_base + 12, 0x777777777777),
> +		cs_mov48(compute_sr_base + 14, 0x888888888888),
> +		/* ShaderProgramDescriptor 0-3 pointers */
> +		cs_mov48(compute_sr_base + 16, spd_gpuva),
> +		cs_mov48(compute_sr_base + 18, 0x999999999999),
> +		cs_mov48(compute_sr_base + 20, 0xaaaaaaaaaaaa),
> +		cs_mov48(compute_sr_base + 22, 0xbbbbbbbbbbbb),
> +		/* ThreadStorageDescriptor 0-3 pointers */
> +		cs_mov48(compute_sr_base + 24, tsd_gpuva),
> +		cs_mov48(compute_sr_base + 26, 0xdddddddddddd),
> +		cs_mov48(compute_sr_base + 28, 0xeeeeeeeeeeee),
> +		cs_mov48(compute_sr_base + 30, 0xffffffffffff),
> +		/* global ID offset = 0 */
> +		cs_mov32(compute_sr_base + 32, 0),
> +		/* workgroup size 64x1x1, allow merge */
> +		cs_mov32(compute_sr_base + 33, (63 << 0) | (0 << 10) | (0 << 20) | (1 << 31)),
> +		/* job offset (x/y/z) = { 0, 0, 0 } */
> +		cs_mov32(compute_sr_base + 34, 0),
> +		cs_mov32(compute_sr_base + 35, 0),
> +		cs_mov32(compute_sr_base + 36, 0),
> +		/* job size (x/y/z) = { 64, 1, 1 } */
> +		cs_mov32(compute_sr_base + 37, 64),
> +		cs_mov32(compute_sr_base + 38, 1),
> +		cs_mov32(compute_sr_base + 39, 1),
> +		/* SET_SB_ENTRY.async_ep1.other1 */
> +		cs_set_sb_entry(1, 1),
> +		/* RUN_COMPUTE.increment64.x_axis.no_increment.srt0.spd0.tsd0.fau0 */
> +		cs_run_compute(64, CS_TASK_AXIS_X, 0, 0, 0, 0, 0),
> +		/* WAIT (RUN_COMPUTE) */
> +		cs_wait(0xff, false),
> +		/* FLUSH_CACHE.clean_invalidate_l2.clean_invalidate_lsc.invalidate_other.sb0 */
> +		cs_flush(CS_FLUSH_MODE_CLEAN_AND_INVALIDATE,
> +			 CS_FLUSH_MODE_CLEAN_AND_INVALIDATE,
> +			 CS_FLUSH_MODE_INVALIDATE,
> +			 0, zero_reg, 1),
> +		/* WAIT (flush) */
> +		cs_wait(0xff, false),
> +		/* ERROR_BARRIER */
> +		cs_error_barrier(),
> +	};
> +
> +	memcpy(gpu_cmdbuf, gpu_instrs, sizeof(gpu_instrs));
> +	*gpu_cmdbuf_size = sizeof(gpu_instrs);
> +	memcpy(csf_cmdbuf, csf_instrs, sizeof(csf_instrs));
> +	*csf_cmdbuf_size = sizeof(csf_instrs);
> +}
> +
>  static size_t
>  issue_store_multiple(uint8_t *cs, uint64_t kernel_va, uint32_t constant)
>  {
> @@ -118,6 +245,119 @@ int igt_main() {
>  		igt_panthor_free_bo(fd, &result_bo);
>  	}
>  
> +	igt_describe("Submit a simple compute job with memory access.");
> +	igt_subtest("simple_compute") {
> +		uint32_t vm_id;
> +		uint32_t group_handle;
> +		struct panthor_bo csf_cmdbuf_bo = {};
> +		struct panthor_bo gpu_cmdbuf_bo = {};
> +		struct panthor_bo resource_table_bo = {};
> +		struct panthor_bo data_bo = {};
> +		uint64_t csf_cmdbuf_gpuva;
> +		uint32_t csf_cmdbuf_size;
> +		uint64_t gpu_cmdbuf_gpuva;
> +		uint32_t gpu_cmdbuf_size;
> +		uint32_t resource_table_gpuva;
> +		uint32_t data_gpuva;
> +		uint32_t syncobj_handle;
> +		uint32_t srt_offset;
> +		uint32_t fau_offset;
> +		uint32_t spd_offset;
> +		uint32_t *spd_ptr;
> +		uint32_t spd_gpuva;
> +		uint32_t tsd_offset;
> +		uint32_t *tsd_ptr;
> +		int arch = igt_panthor_get_arch(fd);
> +
> +		igt_require(arch < 15);
> +
> +		igt_panthor_vm_create(fd, &vm_id, 0);
> +
> +		/* CSF ISA command stream */
> +		csf_cmdbuf_gpuva = 0x1000000;
> +		igt_panthor_bo_create_mapped(fd, &csf_cmdbuf_bo, 4096, 0, 0);
> +		igt_panthor_vm_bind(fd, vm_id, csf_cmdbuf_bo.handle, csf_cmdbuf_gpuva,
> +				    csf_cmdbuf_bo.size, DRM_PANTHOR_VM_BIND_OP_TYPE_MAP, 0);
> +		memset(csf_cmdbuf_bo.map, 0, csf_cmdbuf_bo.size);
> +
> +		/* GPU ISA command stream */
> +		gpu_cmdbuf_gpuva = 0x2000000;
> +		igt_panthor_bo_create_mapped(fd, &gpu_cmdbuf_bo, 4096, 0, 0);
> +		igt_panthor_vm_bind(fd, vm_id, gpu_cmdbuf_bo.handle, gpu_cmdbuf_gpuva,
> +				    gpu_cmdbuf_bo.size, DRM_PANTHOR_VM_BIND_OP_TYPE_MAP, 0);
> +		memset(gpu_cmdbuf_bo.map, 0, gpu_cmdbuf_bo.size);
> +
> +		/* resource tables passed to RUN_COMPUTE */
> +		resource_table_gpuva = 0x3000000;
> +		igt_panthor_bo_create_mapped(fd, &resource_table_bo, 4096, 0, 0);
> +		igt_panthor_vm_bind(fd, vm_id, resource_table_bo.handle, resource_table_gpuva,
> +				    resource_table_bo.size, DRM_PANTHOR_VM_BIND_OP_TYPE_MAP, 0);
> +		memset(resource_table_bo.map, 0, resource_table_bo.size);
> +		/* ShaderResourceTable descriptor (zeroed) */
> +		srt_offset = 0;
> +		/* FastAccessUniform descriptor (zeroed) */
> +		fau_offset = 128;
> +		/* ShaderProgramDescriptor */
> +		spd_offset = 256;
> +		spd_gpuva = resource_table_gpuva + spd_offset;
> +		if (arch == 14)
> +			spd_gpuva |= 2; /* register allocation propagated into ptr */
> +		spd_ptr = (uint32_t *) ((unsigned char *) resource_table_bo.map + spd_offset);
> +		spd_ptr[0] = 0x8 | (0x1 << 4); /* compute shader */
> +		spd_ptr[0] |= (0 << 16) | (0 << 17) | (0 << 19); /* no NaN suppress, FTZ, inf suppress */
> +		spd_ptr[0] |= (0 << 29) | (2 << 30); /* no JUMP_EX, max 32 registers */
> +		spd_ptr[1] |= (1 << 12); /* preload GlobalInvocationID.x into r60 */
> +		spd_ptr[2] = gpu_cmdbuf_gpuva; /* initial PC */
> +		/* ThreadStorageDescriptor */
> +		tsd_offset = 512;
> +		tsd_ptr = (uint32_t *) ((unsigned char *) resource_table_bo.map + tsd_offset);
> +		tsd_ptr[1] = 0x1f; /* max wls_instances */
> +
> +		/* CPU<->GPU shared data */
> +		data_gpuva = 0x4000000;
> +		igt_panthor_bo_create_mapped(fd, &data_bo, 16384, 0, 0);
> +		igt_panthor_vm_bind(fd, vm_id, data_bo.handle, data_gpuva,
> +				    data_bo.size, DRM_PANTHOR_VM_BIND_OP_TYPE_MAP, 0);
> +		for (unsigned int i = 0; i < 64 * 64; i++) {
> +			uint32_t *x = data_bo.map + (i * sizeof(uint32_t));
> +			*x = i;
> +		}
> +
> +		record_compute_prog(csf_cmdbuf_bo.map, &csf_cmdbuf_size,
> +				    gpu_cmdbuf_bo.map, &gpu_cmdbuf_size,
> +				    resource_table_gpuva + srt_offset,
> +				    resource_table_gpuva + fau_offset,
> +				    spd_gpuva,
> +				    resource_table_gpuva + tsd_offset,
> +				    data_gpuva);
> +
> +		group_handle = igt_panthor_group_create_simple(fd, vm_id, 0);
> +		igt_assert_neq(group_handle, 0);
> +
> +		syncobj_handle = syncobj_create(fd, 0);
> +
> +		igt_panthor_group_submit_simple(fd, group_handle, 0,
> +						csf_cmdbuf_gpuva, csf_cmdbuf_size,
> +						syncobj_handle, 0);
> +		igt_assert(syncobj_wait(fd, &syncobj_handle, 1, INT64_MAX, 0, NULL));
> +
> +		for (unsigned int i = 0; i < 64 * 64; i++) {
> +			uint32_t *x = data_bo.map + (i * sizeof(uint32_t));
> +			igt_assert_eq(*x, i * i);
> +		}
> +
> +		syncobj_destroy(fd, syncobj_handle);
> +
> +		igt_panthor_group_destroy(fd, group_handle, 0);
> +
> +		igt_panthor_vm_destroy(fd, vm_id, 0);
> +
> +		igt_panthor_free_bo(fd, &data_bo);
> +		igt_panthor_free_bo(fd, &resource_table_bo);
> +		igt_panthor_free_bo(fd, &gpu_cmdbuf_bo);
> +		igt_panthor_free_bo(fd, &csf_cmdbuf_bo);
> +	}
> +
>  	igt_fixture() {
>  		drm_close_driver(fd);
>  	}
> -- 
> 2.47.3
> 

-- 
====================
| I would like to |
| fix the world,  |
| but they're not |
| giving me the   |
 \ source code!  /
  ---------------
    ¯\_(ツ)_/¯