Skip to contentSkip to conversation

Optimized division operation to shift operation

PATCH Applied Bernard Zhao 7 participants 7 messages Apr 14, 2020 – Apr 15, 2020

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Optimized division operation to shift operation

On some processors, the / operate will call the compiler`s div lib, which is low efficient, We can replace the / operation with shift, so that we can replace the call of the division library with one shift assembly instruction.

---
 drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c | 4 ++--
 drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c | 4 ++--
 drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c | 4 ++--
 3 files changed, 6 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c +2 -2
index b205039..66cd078 100644
@@ -175,10 +175,10 @@ static int gmc_v6_0_mc_load_microcode(struct amdgpu_device *adev)
 	amdgpu_ucode_print_mc_hdr(&hdr->header);
 
 	adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-	regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+	regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
 	new_io_mc_regs = (const __le32 *)
 		(adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-	ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+	ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
 	new_fw_data = (const __le32 *)
 		(adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c +2 -2
index 9da9596..ca26d63 100644
@@ -193,10 +193,10 @@ static int gmc_v7_0_mc_load_microcode(struct amdgpu_device *adev)
 	amdgpu_ucode_print_mc_hdr(&hdr->header);
 
 	adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-	regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+	regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
 	io_mc_regs = (const __le32 *)
 		(adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-	ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+	ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
 	fw_data = (const __le32 *)
 		(adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c +2 -2
index 27d83204..295039c 100644
@@ -318,10 +318,10 @@ static int gmc_v8_0_tonga_mc_load_microcode(struct amdgpu_device *adev)
 	amdgpu_ucode_print_mc_hdr(&hdr->header);
 
 	adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-	regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+	regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
 	io_mc_regs = (const __le32 *)
 		(adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-	ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+	ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
 	fw_data = (const __le32 *)
 		(adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
Signature · 1 line
2.7.4
Applied
#1 lore raw
Quoted 7 lines · Bernard Zhao

On some processors, the / operate will call the compiler`s div lib, which is low efficient, We can replace the / operation with shift, so that we can replace the call of the division library with one shift assembly instruction.

Applied. thanks.

Alex

Quoted 60 lines
---
 drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c | 4 ++--
 drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c | 4 ++--
 drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c | 4 ++--
 3 files changed, 6 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c +2 -2
index b205039..66cd078 100644
@@ -175,10 +175,10 @@ static int gmc_v6_0_mc_load_microcode(struct amdgpu_device *adev)
        amdgpu_ucode_print_mc_hdr(&hdr->header);
 
        adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
        new_io_mc_regs = (const __le32 *)
                (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
        new_fw_data = (const __le32 *)
                (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c +2 -2
index 9da9596..ca26d63 100644
@@ -193,10 +193,10 @@ static int gmc_v7_0_mc_load_microcode(struct amdgpu_device *adev)
        amdgpu_ucode_print_mc_hdr(&hdr->header);
 
        adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
        io_mc_regs = (const __le32 *)
                (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
        fw_data = (const __le32 *)
                (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c +2 -3
index 27d83204..295039c 100644
@@ -318,10 +318,10 @@ static int gmc_v8_0_tonga_mc_load_microcode(struct amdgpu_device *adev)
        amdgpu_ucode_print_mc_hdr(&hdr->header);
 
        adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
        io_mc_regs = (const __le32 *)
                (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
        fw_data = (const __le32 *)
                (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
--

2.7.4

Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#2 lore raw
Quoted 6 lines · Alex Deucher
Quoted 5 lines · Bernard Zhao

On some processors, the / operate will call the compiler`s div lib, which is low efficient, We can replace the / operation with shift, so that we can replace the call of the division library with one shift assembly instruction.

This was applied already, and it's not in a driver I look after... but to me this feels like something that really should be justified. Using >> instead of / for multiples of 2 division mattered 20 years ago, I'd be surprised if it still did on modern compilers.

BR,
Jani.

Quoted 75 lines
Quoted 2 lines

Applied. thanks.

Alex

Quoted 60 lines
---
 drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c | 4 ++--
 drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c | 4 ++--
 drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c | 4 ++--
 3 files changed, 6 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c +2 -2
index b205039..66cd078 100644
@@ -175,10 +175,10 @@ static int gmc_v6_0_mc_load_microcode(struct amdgpu_device *adev)
        amdgpu_ucode_print_mc_hdr(&hdr->header);
 
        adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
        new_io_mc_regs = (const __le32 *)
                (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
        new_fw_data = (const __le32 *)
                (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c +2 -2
index 9da9596..ca26d63 100644
@@ -193,10 +193,10 @@ static int gmc_v7_0_mc_load_microcode(struct amdgpu_device *adev)
        amdgpu_ucode_print_mc_hdr(&hdr->header);
 
        adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
        io_mc_regs = (const __le32 *)
                (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
        fw_data = (const __le32 *)
                (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c +2 -3
index 27d83204..295039c 100644
@@ -318,10 +318,10 @@ static int gmc_v8_0_tonga_mc_load_microcode(struct amdgpu_device *adev)
        amdgpu_ucode_print_mc_hdr(&hdr->header);
 
        adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
        io_mc_regs = (const __le32 *)
                (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
        fw_data = (const __le32 *)
                (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
--

2.7.4

Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Signature · 1 line
Jani Nikula, Intel Open Source Graphics Center
maintainer
#3 lore raw

Am 15.04.20 um 09:41 schrieb Jani Nikula:

Quoted 10 lines · Alex Deucher
Quoted 5 lines · Bernard Zhao
Quoted 4 lines

On some processors, the / operate will call the compiler`s div lib, which is low efficient, We can replace the / operation with shift, so that we can replace the call of the division library with one shift assembly instruction.

This was applied already, and it's not in a driver I look after... but to me this feels like something that really should be justified. Using >> instead of / for multiples of 2 division mattered 20 years ago, I'd be surprised if it still did on modern compilers.

I have similar worries, especially since we replace the "/ (4 * 2)" with ">> 3" it's making the code just a bit less readable.

And that the code runs exactly once while loading the driver and pushing the firmware into the hardware. So performance is completely irrelevant here.

Regards,
Christian.

Quoted 78 lines

BR,
Jani.

Quoted 73 lines
Quoted 1 line

Applied. thanks.

Alex

Quoted 60 lines
---
  drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c | 4 ++--
  drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c | 4 ++--
  drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c | 4 ++--
  3 files changed, 6 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c +2 -2
index b205039..66cd078 100644
@@ -175,10 +175,10 @@ static int gmc_v6_0_mc_load_microcode(struct amdgpu_device *adev)
         amdgpu_ucode_print_mc_hdr(&hdr->header);
 
         adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
         new_io_mc_regs = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
         new_fw_data = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c +2 -2
index 9da9596..ca26d63 100644
@@ -193,10 +193,10 @@ static int gmc_v7_0_mc_load_microcode(struct amdgpu_device *adev)
         amdgpu_ucode_print_mc_hdr(&hdr->header);
 
         adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
         io_mc_regs = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
         fw_data = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c +2 -3
index 27d83204..295039c 100644
@@ -318,10 +318,10 @@ static int gmc_v8_0_tonga_mc_load_microcode(struct amdgpu_device *adev)
         amdgpu_ucode_print_mc_hdr(&hdr->header);
 
         adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
         io_mc_regs = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
         fw_data = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
--

2.7.4

Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&data=02%7C01%7Cchristian.koenig%40amd.com%7C1e91f7edcfe0473b0d7008d7e11074a8%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637225333103893889&sdata=VDJlEY2%2Bl1SSO8Fw1dYqqPFqQtyHpsxQ0Tm7iVOgJQY%3D&reserved=0

_______________________________________________
dri-devel mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Fdri-devel&data=02%7C01%7Cchristian.koenig%40amd.com%7C1e91f7edcfe0473b0d7008d7e11074a8%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637225333103893889&sdata=EpqRRbCiksur%2BjMlVQplExuJsmw6UPODhyBOutOVukw%3D&reserved=0

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#4 lore raw
Quoted 13 lines · Christian König

Am 15.04.20 um 09:41 schrieb Jani Nikula:

Quoted 10 lines · Alex Deucher
Quoted 5 lines · Bernard Zhao
Quoted 4 lines

On some processors, the / operate will call the compiler`s div lib, which is low efficient, We can replace the / operation with shift, so that we can replace the call of the division library with one shift assembly instruction.

This was applied already, and it's not in a driver I look after... but to me this feels like something that really should be justified. Using >> instead of / for multiples of 2 division mattered 20 years ago, I'd be surprised if it still did on modern compilers.

Quoted 6 lines

I have similar worries, especially since we replace the "/ (4 * 2)" with ">> 3" it's making the code just a bit less readable.

And that the code runs exactly once while loading the driver and pushing the firmware into the hardware. So performance is completely irrelevant here.

Yeah, and even in general I'd really want to see proof that such bad compilers exist first. Doing a peephole pass for stuff like this is like undergrad compiler course stuff. Also I'd trust the compiler's call on which is faster much more than humans doing a manual peephole pass.

Or am I just massively biased from all the layman following of what kind of our gl/vk/compute compiler people are doing on a daily basis? -Daniel

Quoted 87 lines

Regards,
Christian.

Quoted 78 lines

BR,
Jani.

Quoted 73 lines
Quoted 1 line

Applied. thanks.

Alex

Quoted 60 lines
---
  drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c | 4 ++--
  drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c | 4 ++--
  drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c | 4 ++--
  3 files changed, 6 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c +2 -2
index b205039..66cd078 100644
@@ -175,10 +175,10 @@ static int gmc_v6_0_mc_load_microcode(struct amdgpu_device *adev)
         amdgpu_ucode_print_mc_hdr(&hdr->header);
 
         adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
         new_io_mc_regs = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
         new_fw_data = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c +2 -2
index 9da9596..ca26d63 100644
@@ -193,10 +193,10 @@ static int gmc_v7_0_mc_load_microcode(struct amdgpu_device *adev)
         amdgpu_ucode_print_mc_hdr(&hdr->header);
 
         adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
         io_mc_regs = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
         fw_data = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c +2 -3
index 27d83204..295039c 100644
@@ -318,10 +318,10 @@ static int gmc_v8_0_tonga_mc_load_microcode(struct amdgpu_device *adev)
         amdgpu_ucode_print_mc_hdr(&hdr->header);
 
         adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
         io_mc_regs = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
         fw_data = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
--

2.7.4

Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&data=02%7C01%7Cchristian.koenig%40amd.com%7C1e91f7edcfe0473b0d7008d7e11074a8%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637225333103893889&sdata=VDJlEY2%2Bl1SSO8Fw1dYqqPFqQtyHpsxQ0Tm7iVOgJQY%3D&reserved=0

_______________________________________________
dri-devel mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Fdri-devel&data=02%7C01%7Cchristian.koenig%40amd.com%7C1e91f7edcfe0473b0d7008d7e11074a8%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637225333103893889&sdata=EpqRRbCiksur%2BjMlVQplExuJsmw6UPODhyBOutOVukw%3D&reserved=0

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Signature · 7 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#5 lore raw
In reply to Christian König

From: Christian König

Quoted 13 lines

Sent: 15 April 2020 08:57
Am 15.04.20 um 09:41 schrieb Jani Nikula:

Quoted 10 lines · Alex Deucher
Quoted 5 lines · Bernard Zhao
Quoted 4 lines

On some processors, the / operate will call the compiler`s div lib, which is low efficient, We can replace the / operation with shift, so that we can replace the call of the division library with one shift assembly instruction.

This was applied already, and it's not in a driver I look after... but to me this feels like something that really should be justified. Using >> instead of / for multiples of 2 division mattered 20 years ago, I'd be surprised if it still did on modern compilers.

Quoted 6 lines

I have similar worries, especially since we replace the "/ (4 * 2)" with ">> 3" it's making the code just a bit less readable.

And that the code runs exactly once while loading the driver and pushing the firmware into the hardware. So performance is completely irrelevant here.

Force the division to be unsigned and the compiler will use a shift.

David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)

#6 lore raw
In reply to Christian König

[AMD Public Use]

I've gone ahead and dropped the patch.

Alex
________________________________
From: Koenig, Christian <[email protected]>
Sent: Wednesday, April 15, 2020 3:57 AM
To: Jani Nikula <[email protected]>; Alex Deucher <[email protected]>; Bernard Zhao <[email protected]>
Cc: Sierra Guiza, Alejandro (Alex) <[email protected]>; Zeng, Oak <[email protected]>; Maling list - DRI developers <[email protected]>; David Airlie <[email protected]>; Kuehling, Felix <[email protected]>; LKML <[email protected]>; amd-gfx list <[email protected]>; [email protected] <[email protected]>; Huang, Ray <[email protected]>; Russell, Kent <[email protected]>; Deucher, Alexander <[email protected]>; Sam Ravnborg <[email protected]>; Yuan, Xiaojie <[email protected]>
Subject: Re: [PATCH] Optimized division operation to shift operation

Am 15.04.20 um 09:41 schrieb Jani Nikula:

Quoted 10 lines · Alex Deucher
Quoted 5 lines · Bernard Zhao
Quoted 4 lines

On some processors, the / operate will call the compiler`s div lib, which is low efficient, We can replace the / operation with shift, so that we can replace the call of the division library with one shift assembly instruction.

This was applied already, and it's not in a driver I look after... but to me this feels like something that really should be justified. Using >> instead of / for multiples of 2 division mattered 20 years ago, I'd be surprised if it still did on modern compilers.

I have similar worries, especially since we replace the "/ (4 * 2)" with ">> 3" it's making the code just a bit less readable.

And that the code runs exactly once while loading the driver and pushing the firmware into the hardware. So performance is completely irrelevant here.

Regards,
Christian.

Quoted 78 lines

BR,
Jani.

Quoted 73 lines
Quoted 1 line

Applied. thanks.

Alex

Quoted 60 lines
---
  drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c | 4 ++--
  drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c | 4 ++--
  drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c | 4 ++--
  3 files changed, 6 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c +2 -2
index b205039..66cd078 100644
@@ -175,10 +175,10 @@ static int gmc_v6_0_mc_load_microcode(struct amdgpu_device *adev)
         amdgpu_ucode_print_mc_hdr(&hdr->header);
 
         adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
         new_io_mc_regs = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
         new_fw_data = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c +2 -2
index 9da9596..ca26d63 100644
@@ -193,10 +193,10 @@ static int gmc_v7_0_mc_load_microcode(struct amdgpu_device *adev)
         amdgpu_ucode_print_mc_hdr(&hdr->header);
 
         adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
         io_mc_regs = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
         fw_data = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
drivers/gpu/drm/amd/amdgpu/gmc_v8_0.c +2 -3
index 27d83204..295039c 100644
@@ -318,10 +318,10 @@ static int gmc_v8_0_tonga_mc_load_microcode(struct amdgpu_device *adev)
         amdgpu_ucode_print_mc_hdr(&hdr->header);
 
         adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
-       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
+       regs_size = le32_to_cpu(hdr->io_debug_size_bytes) >> 3;
         io_mc_regs = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
-       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
+       ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) >> 2;
         fw_data = (const __le32 *)
                 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 
--

2.7.4

Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Cchristian.koenig%40amd.com%7C1e91f7edcfe0473b0d7008d7e11074a8%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637225333103893889&amp;sdata=VDJlEY2%2Bl1SSO8Fw1dYqqPFqQtyHpsxQ0Tm7iVOgJQY%3D&amp;reserved=0

_______________________________________________
dri-devel mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Fdri-devel&amp;data=02%7C01%7Cchristian.koenig%40amd.com%7C1e91f7edcfe0473b0d7008d7e11074a8%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637225333103893889&amp;sdata=EpqRRbCiksur%2BjMlVQplExuJsmw6UPODhyBOutOVukw%3D&amp;reserved=0

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx