Skip to contentSkip to conversation

More sm501 fixes and optimisations

PATCH Superseded BALATON Zoltan 2 participants 15 messages Jun 6, 2020 – Jun 15, 2020

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

cover More sm501 fixes and optimisations

Some small fixes and optimisations for hw/display/sm501.c.

BALATON Zoltan (4):
sm501: Fix bounds checks
sm501: Drop unneded variable
sm501: Ignore no-op blits
sm501: Optimise 1 pixel 2d ops

 hw/display/sm501.c | 58 ++++++++++++++++++++++++++++++++++++++--------
 1 file changed, 48 insertions(+), 10 deletions(-)
Signature · 1 line
2.21.3
#1 lore raw

PATCH 1/4 sm501: Fix bounds checks

We don't need to add width to pitch when calculating last point, that would reject valid ops within the card's local_mem.

Fixes b15a22bbcbe6a78dc3d88fe3134985e4cdd87de4
---
 hw/display/sm501.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
hw/display/sm501.c +4 -4
index edd8d24a76..5ae320ddc3 100644
@@ -723,8 +723,8 @@ static void sm501_2d_operation(SM501State *s)
         dst_y -= height - 1;
     }
 
-    if (dst_base >= get_local_mem_size(s) || dst_base +
-        (dst_x + width + (dst_y + height) * (dst_pitch + width)) *
+    if (dst_base >= get_local_mem_size(s) ||
+        dst_base + (dst_x + width + (dst_y + height) * dst_pitch) *
         (1 << format) >= get_local_mem_size(s)) {
         qemu_log_mask(LOG_GUEST_ERROR, "sm501: 2D op dest is outside vram.\n");
         return;
@@ -749,8 +749,8 @@ static void sm501_2d_operation(SM501State *s)
             src_y -= height - 1;
         }
 
-        if (src_base >= get_local_mem_size(s) || src_base +
-            (src_x + width + (src_y + height) * (src_pitch + width)) *
+        if (src_base >= get_local_mem_size(s) ||
+            src_base + (src_x + width + (src_y + height) * src_pitch) *
             (1 << format) >= get_local_mem_size(s)) {
             qemu_log_mask(LOG_GUEST_ERROR,
                           "sm501: 2D op src is outside vram.\n");
Signature · 1 line
2.21.3
Reviewed-by on v1 1/4
#2 lore raw

Re: [PATCH 1/4] sm501: Fix bounds checks

Quoted 9 lines · BALATON Zoltan

We don't need to add width to pitch when calculating last point, that would reject valid ops within the card's local_mem.

Fixes b15a22bbcbe6a78dc3d88fe3134985e4cdd87de4
---
 hw/display/sm501.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

The calculations for sb/se/db/de all have a term which multiplies by (width + pitch), which makes me suspect they also need a similar fix ?

thanks
-- PMM

#3 lore raw

Re: [PATCH 1/4] sm501: Fix bounds checks

Quoted 11 lines · Peter Maydell
Quoted 9 lines · BALATON Zoltan

We don't need to add width to pitch when calculating last point, that would reject valid ops within the card's local_mem.

Fixes b15a22bbcbe6a78dc3d88fe3134985e4cdd87de4
---
 hw/display/sm501.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
Quoted 5 lines

The calculations for sb/se/db/de all have a term which multiplies by (width + pitch), which makes me suspect they also need a similar fix ?

Maybe. I'll have to check again. Actually is there a simpler way to check if two rectangles overlap when they are given with base, x, y, w, h, pitch where base is the first byte of the screen, pitch is length of one line and x,y is coordinates of top left corner and w,h is dimensions of the rect. Now that I think about it we also need to take into accounf the bytes per pixel value (1 << format) because base is given in bytes while others are in pixels so these formulae likely need some fixes. Pixman has some functions for these but those assume common base so to use those we would need to bring the two rectangles to common base which I could not find out how to do. Probably this is really simple for someone who already did a lot of these before.

Regards,
BALATON Zoltan

#4 lore raw

Re: [PATCH 1/4] sm501: Fix bounds checks

Quoted 6 lines · BALATON Zoltan
Quoted 3 lines · Peter Maydell

The calculations for sb/se/db/de all have a term which multiplies by (width + pitch), which makes me suspect they also need a similar fix ?

Quoted 11 lines

Maybe. I'll have to check again. Actually is there a simpler way to check if two rectangles overlap when they are given with base, x, y, w, h, pitch where base is the first byte of the screen, pitch is length of one line and x,y is coordinates of top left corner and w,h is dimensions of the rect. Now that I think about it we also need to take into accounf the bytes per pixel value (1 << format) because base is given in bytes while others are in pixels so these formulae likely need some fixes. Pixman has some functions for these but those assume common base so to use those we would need to bring the two rectangles to common base which I could not find out how to do. Probably this is really simple for someone who already did a lot of these before.

I think the thing that makes it particularly awkward is that the source and dest can have different pitches. That means it's not a simple "do two rectangles overlap" test because the dest area might not be a rectangle at all when looked at from the POV of the source.

Do guests usually set src and dst pitch identical? If so it might be worth having a more accurate rectangle-overlap test for the common case and a looser check for the hard-to-handle case.

I might have a think about this and draw some diagrams tomorrow :-)

thanks
-- PMM

#5 lore raw

Re: [PATCH 1/4] sm501: Fix bounds checks

Quoted 19 lines · Peter Maydell
Quoted 6 lines · BALATON Zoltan
Quoted 3 lines · Peter Maydell

The calculations for sb/se/db/de all have a term which multiplies by (width + pitch), which makes me suspect they also need a similar fix ?

Quoted 11 lines

Maybe. I'll have to check again. Actually is there a simpler way to check if two rectangles overlap when they are given with base, x, y, w, h, pitch where base is the first byte of the screen, pitch is length of one line and x,y is coordinates of top left corner and w,h is dimensions of the rect. Now that I think about it we also need to take into accounf the bytes per pixel value (1 << format) because base is given in bytes while others are in pixels so these formulae likely need some fixes. Pixman has some functions for these but those assume common base so to use those we would need to bring the two rectangles to common base which I could not find out how to do. Probably this is really simple for someone who already did a lot of these before.

Quoted 2 lines

I think the thing that makes it particularly awkward is that the source and dest can have different pitches. That means it's

Yes, different bases and different pitches make it difficult to compare and also to transform them to a common base to be able to do rectangle overlap test (I haven't checked but maybe the pitches could be relative prime so it's possible there's no common pitch to transform them to a single view).

Quoted 8 lines

not a simple "do two rectangles overlap" test because the dest area might not be a rectangle at all when looked at from the POV of the source.

Do guests usually set src and dst pitch identical? If so it might be worth having a more accurate rectangle-overlap test for the common case and a looser check for the hard-to-handle case.

Quick test (just booted AmigaOS, opened some windows and moved them around a bit) shows same pitch is common but not always:

7610 sm501 2d engine regs : write addr=10, val=200020 4249 sm501 2d engine regs : write addr=10, val=4000400 2340 sm501 2d engine regs : write addr=10, val=4000080

    462 sm501 2d engine regs : write addr=10, val=40000a0
    453 sm501 2d engine regs : write addr=10, val=800080
    434 sm501 2d engine regs : write addr=10, val=1200120
    233 sm501 2d engine regs : write addr=10, val=4000020
    162 sm501 2d engine regs : write addr=10, val=400040
    144 sm501 2d engine regs : write addr=10, val=200080
    141 sm501 2d engine regs : write addr=10, val=2000a0
     75 sm501 2d engine regs : write addr=10, val=600060
     62 sm501 2d engine regs : write addr=10, val=a000a0
     62 sm501 2d engine regs : write addr=10, val=200400
     38 sm501 2d engine regs : write addr=10, val=e000e0
     38 sm501 2d engine regs : write addr=10, val=1000100
     38 sm501 2d engine regs : write addr=10, val=10000a0
     37 sm501 2d engine regs : write addr=10, val=4000040
     36 sm501 2d engine regs : write addr=10, val=40000e0
     34 sm501 2d engine regs : write addr=10, val=a00080
     34 sm501 2d engine regs : write addr=10, val=4000120
     33 sm501 2d engine regs : write addr=10, val=400080
     25 sm501 2d engine regs : write addr=10, val=600080
     23 sm501 2d engine regs : write addr=10, val=4000060
     22 sm501 2d engine regs : write addr=10, val=1a001a0

and may depend on what the guest is doing so not sure it's worth the effort. (Probably same pitch is used when moving window and different when drawing something like text or icons but different pitch likely uses off-screen bitmap as source so overlap is not likely for those.)

Quoted 1 line

I might have a think about this and draw some diagrams tomorrow :-)

I'll wait for that to see if you can come up with something clever. I've tried to find a solution but could not come up with anything simpler than a loop checking lines which is probably as much effort as doing the blit itself and then doing blit with a temporary is probably the same so I went for just checking the obviously far apart blits (and even that seems to be easy to miss). So I'm happy if we can just fix the checks we have but if there are better suggestions please tell me.

Regards,
BALATON Zoltan

#6 lore raw

PATCH 2/4 sm501: Drop unneded variable

In reply to BALATON Zoltan

We don't need a separate variable to keep track if we allocated memory that needs to be freed as we can test the pointer itself.

---
 hw/display/sm501.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)
hw/display/sm501.c +2 -3
index 5ae320ddc3..85d54b598f 100644
@@ -796,13 +796,12 @@ static void sm501_2d_operation(SM501State *s)
             de = db + width + height * (width + dst_pitch);
             if (rtl && ((db >= sb && db <= se) || (de >= sb && de <= se))) {
                 /* regions may overlap: copy via temporary */
-                int free_buf = 0, llb = width * (1 << format);
+                int llb = width * (1 << format);
                 int tmp_stride = DIV_ROUND_UP(llb, sizeof(uint32_t));
                 uint32_t *tmp = tmp_buf;
 
                 if (tmp_stride * sizeof(uint32_t) * height > sizeof(tmp_buf)) {
                     tmp = g_malloc(tmp_stride * sizeof(uint32_t) * height);
-                    free_buf = 1;
                 }
                 pixman_blt((uint32_t *)&s->local_mem[src_base], tmp,
                            src_pitch * (1 << format) / sizeof(uint32_t),
@@ -813,7 +812,7 @@ static void sm501_2d_operation(SM501State *s)
                            dst_pitch * (1 << format) / sizeof(uint32_t),
                            8 * (1 << format), 8 * (1 << format),
                            0, 0, dst_x, dst_y, width, height);
-                if (free_buf) {
+                if (tmp != tmp_buf) {
                     g_free(tmp);
                 }
             } else {
Signature · 1 line
2.21.3
Reviewed-by on v1 2/4
#7 lore raw

Re: [PATCH 2/4] sm501: Drop unneded variable

Quoted 8 lines · BALATON Zoltan

We don't need a separate variable to keep track if we allocated memory that needs to be freed as we can test the pointer itself.

---
 hw/display/sm501.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

thanks
-- PMM

#8 lore raw

PATCH 3/4 sm501: Ignore no-op blits

In reply to BALATON Zoltan

Some guests seem to try source copy blits with same source and dest which are no-op so avoid calling pixman for these.

---
 hw/display/sm501.c | 5 +++++
 1 file changed, 5 insertions(+)
hw/display/sm501.c +5 -0
index 85d54b598f..3397ca9fbf 100644
@@ -788,6 +788,11 @@ static void sm501_2d_operation(SM501State *s)
                               (rop2_source_is_pattern ?
                                   " with pattern source" : ""));
             }
+            /* Ignore no-op blits, some guests seem to do this */
+            if (src_base == dst_base && src_pitch == dst_pitch &&
+                src_x == dst_x && src_y == dst_y) {
+                break;
+            }
             /* Check for overlaps, this could be made more exact */
             uint32_t sb, se, db, de;
             sb = src_base + src_x + src_y * (width + src_pitch);
Signature · 1 line
2.21.3
Reviewed-by on v1 3/4
#9 lore raw

Re: [PATCH 3/4] sm501: Ignore no-op blits

Quoted 3 lines · BALATON Zoltan

Some guests seem to try source copy blits with same source and dest which are no-op so avoid calling pixman for these.

Are they doing actual source copy blits, or one of the other currently-unimplemented op types which we currently fall back to doing as source-copy ?

Quoted 1 line

thanks
-- PMM

#10 lore raw

Re: [PATCH 3/4] sm501: Ignore no-op blits

Quoted 8 lines · Peter Maydell
Quoted 3 lines · BALATON Zoltan

Some guests seem to try source copy blits with same source and dest which are no-op so avoid calling pixman for these.

Are they doing actual source copy blits, or one of the other currently-unimplemented op types which we currently fall back to doing as source-copy ?

I thought the same but in the cases I've seen the rop was 0xcc which is copy source so it looks like an actual source copy blit (this seems to be an issue in the guest driver, I've notified appropriate people but not sure it's still maintained). Even if it was fallback we wouldn't need to do the copy in this case as it changes nothing so we can optimise this out. When we implement the missing op that would come before this case so this is only for source copy case.

Quoted 3 lines
Quoted 1 line

Thank you,
BALATON Zoltan

#11 lore raw

PATCH 4/4 sm501: Optimise 1 pixel 2d ops

In reply to BALATON Zoltan

Some guests do 1x1 blits which is faster to do directly than calling a function for it so avoid overhead in this case.

---
 hw/display/sm501.c | 40 +++++++++++++++++++++++++++++++++++++---
 1 file changed, 37 insertions(+), 3 deletions(-)
hw/display/sm501.c +37 -3
index 3397ca9fbf..59693fbb5c 100644
@@ -793,6 +793,25 @@ static void sm501_2d_operation(SM501State *s)
                 src_x == dst_x && src_y == dst_y) {
                 break;
             }
+            /* Some clients also do 1 pixel blits, avoid overhead for these */
+            if (width == 1 && height == 1) {
+                unsigned int si = (src_x + src_y * src_pitch) * (1 << format);
+                unsigned int di = (dst_x + dst_y * dst_pitch) * (1 << format);
+                switch (format) {
+                case 0:
+                    s->local_mem[dst_base + di] = s->local_mem[src_base + si];
+                    break;
+                case 1:
+                    *(uint16_t *)&s->local_mem[dst_base + di] =
+                                    *(uint16_t *)&s->local_mem[src_base + si];
+                    break;
+                case 2:
+                    *(uint32_t *)&s->local_mem[dst_base + di] =
+                                    *(uint32_t *)&s->local_mem[src_base + si];
+                    break;
+                }
+                break;
+            }
             /* Check for overlaps, this could be made more exact */
             uint32_t sb, se, db, de;
             sb = src_base + src_x + src_y * (width + src_pitch);
@@ -841,9 +860,24 @@ static void sm501_2d_operation(SM501State *s)
             color = cpu_to_le16(color);
         }
 
-        pixman_fill((uint32_t *)&s->local_mem[dst_base],
-                    dst_pitch * (1 << format) / sizeof(uint32_t),
-                    8 * (1 << format), dst_x, dst_y, width, height, color);
+        if (width == 1 && height == 1) {
+            unsigned int i = (dst_x + dst_y * dst_pitch) * (1 << format);
+            switch (format) {
+            case 0:
+                s->local_mem[dst_base + i] = color & 0xff;
+                break;
+            case 1:
+                *(uint16_t *)&s->local_mem[dst_base + i] = color & 0xffff;
+                break;
+            case 2:
+                *(uint32_t *)&s->local_mem[dst_base + i] = color;
+                break;
+            }
+        } else {
+            pixman_fill((uint32_t *)&s->local_mem[dst_base],
+                        dst_pitch * (1 << format) / sizeof(uint32_t),
+                        8 * (1 << format), dst_x, dst_y, width, height, color);
+        }
         break;
     }
     default:
Signature · 1 line
2.21.3
#12 lore raw

Re: [PATCH 4/4] sm501: Optimise 1 pixel 2d ops

Quoted 3 lines · BALATON Zoltan

Some guests do 1x1 blits which is faster to do directly than calling a function for it so avoid overhead in this case.

How much does the performance improve by ?

Quoted 20 lines
---
 hw/display/sm501.c | 40 +++++++++++++++++++++++++++++++++++++---
 1 file changed, 37 insertions(+), 3 deletions(-)
hw/display/sm501.c +7 -0
index 3397ca9fbf..59693fbb5c 100644
@@ -793,6 +793,25 @@ static void sm501_2d_operation(SM501State *s)
                 src_x == dst_x && src_y == dst_y) {
                 break;
             }
+            /* Some clients also do 1 pixel blits, avoid overhead for these */
+            if (width == 1 && height == 1) {
+                unsigned int si = (src_x + src_y * src_pitch) * (1 << format);
+                unsigned int di = (dst_x + dst_y * dst_pitch) * (1 << format);
+                switch (format) {
+                case 0:
+                    s->local_mem[dst_base + di] = s->local_mem[src_base + si];

hw/display/sm501.c static void sm501_2d_operation(SM501State *s) · line 812

Quoted 10 lines
hw/display/sm501.c +10 -0
+                    break;
+                case 1:
+                    *(uint16_t *)&s->local_mem[dst_base + di] =
+                                    *(uint16_t *)&s->local_mem[src_base + si];
+                    break;
+                case 2:
+                    *(uint32_t *)&s->local_mem[dst_base + di] =
+                                    *(uint32_t *)&s->local_mem[src_base + si];
+                    break;
+                }

You could write this more compactly as

                   stn_he_p(&s->local_mem[dst_base + di], 1 << format,
                            ldn_he_p(&s->local_mem[src_base + si], 1

<< format));

(which handles the length-cases for you and also doesn't rely on casting a uint8_t* giving you something correctly aligned for a wider access).

Quoted 15 lines
hw/display/sm501.c +5 -3
+                break;
+            }
             /* Check for overlaps, this could be made more exact */
             uint32_t sb, se, db, de;
             sb = src_base + src_x + src_y * (width + src_pitch);
@@ -841,9 +860,24 @@ static void sm501_2d_operation(SM501State *s)
             color = cpu_to_le16(color);
         }
 
-        pixman_fill((uint32_t *)&s->local_mem[dst_base],
-                    dst_pitch * (1 << format) / sizeof(uint32_t),
-                    8 * (1 << format), dst_x, dst_y, width, height, color);
+        if (width == 1 && height == 1) {
+            unsigned int i = (dst_x + dst_y * dst_pitch) * (1 << format);
+            switch (format) {

hw/display/sm501.c static void sm501_2d_operation(SM501State *s) · line 875

Quoted 10 lines
hw/display/sm501.c +10 -0
+            case 0:
+                s->local_mem[dst_base + i] = color & 0xff;
+                break;
+            case 1:
+                *(uint16_t *)&s->local_mem[dst_base + i] = color & 0xffff;
+                break;
+            case 2:
+                *(uint32_t *)&s->local_mem[dst_base + i] = color;
+                break;
+            }
               stn_he_p(&s->local_mem[dst_base + i], 1 << format, color);

hw/display/sm501.c static void sm501_2d_operation(SM501State *s) · line 883

Quoted 8 lines
hw/display/sm501.c +5 -0
+        } else {
+            pixman_fill((uint32_t *)&s->local_mem[dst_base],
+                        dst_pitch * (1 << format) / sizeof(uint32_t),
+                        8 * (1 << format), dst_x, dst_y, width, height, color);
+        }
         break;
     }
     default:

thanks
-- PMM

#13 lore raw

Re: [PATCH 4/4] sm501: Optimise 1 pixel 2d ops

Quoted 6 lines · Peter Maydell
Quoted 3 lines · BALATON Zoltan

Some guests do 1x1 blits which is faster to do directly than calling a function for it so avoid overhead in this case.

How much does the performance improve by ?

I haven't actually measured due to lack of time experimenting with it and those who I've asked to check it only reported it felt a bit faster but no numerical measurements so that does not prove anything. Anyway these special cases should not hurt and simple enough to have it here even if may not improve performance very much. The biggest loss is probably converting 16 bit screen on every display update to 32 bit because the guest driver does not allow higher bit depth than 16 for some reason. But I don't plan to spend more time with improving sm501, only done it because I had to touch it anyway. Longer term plan is to finish ati-vga which should have better guest drivers and better performance.

Quoted 32 lines
Quoted 20 lines
---
 hw/display/sm501.c | 40 +++++++++++++++++++++++++++++++++++++---
 1 file changed, 37 insertions(+), 3 deletions(-)
hw/display/sm501.c +7 -0
index 3397ca9fbf..59693fbb5c 100644
@@ -793,6 +793,25 @@ static void sm501_2d_operation(SM501State *s)
                 src_x == dst_x && src_y == dst_y) {
                 break;
             }
+            /* Some clients also do 1 pixel blits, avoid overhead for these */
+            if (width == 1 && height == 1) {
+                unsigned int si = (src_x + src_y * src_pitch) * (1 << format);
+                unsigned int di = (dst_x + dst_y * dst_pitch) * (1 << format);
+                switch (format) {
+                case 0:
+                    s->local_mem[dst_base + di] = s->local_mem[src_base + si];

hw/display/sm501.c static void sm501_2d_operation(SM501State *s) · line 812

Quoted 10 lines
hw/display/sm501.c +10 -0
+                    break;
+                case 1:
+                    *(uint16_t *)&s->local_mem[dst_base + di] =
+                                    *(uint16_t *)&s->local_mem[src_base + si];
+                    break;
+                case 2:
+                    *(uint32_t *)&s->local_mem[dst_base + di] =
+                                    *(uint32_t *)&s->local_mem[src_base + si];
+                    break;
+                }

You could write this more compactly as

Quoted 7 lines
                  stn_he_p(&s->local_mem[dst_base + di], 1 << format,
                           ldn_he_p(&s->local_mem[src_base + si], 1

<< format));

(which handles the length-cases for you and also doesn't rely on casting a uint8_t* giving you something correctly aligned for a wider access).

OK, I never know these and they are a bit hard to find because they are defined as glued together macros so grepping for it does not show the definition. Maybe adding a comment with the names where these are defined in bswap.h might help. stn_he_p seems to ultimately call __builtin_memcpy that probably does the same direct assignment for short values. I wonder how much overhead that has going through all the function calls but hopefully those are inlined.

Regards,
BALATON Zoltan

Quoted 27 lines
Quoted 15 lines
hw/display/sm501.c +5 -3
+                break;
+            }
             /* Check for overlaps, this could be made more exact */
             uint32_t sb, se, db, de;
             sb = src_base + src_x + src_y * (width + src_pitch);
@@ -841,9 +860,24 @@ static void sm501_2d_operation(SM501State *s)
             color = cpu_to_le16(color);
         }
 
-        pixman_fill((uint32_t *)&s->local_mem[dst_base],
-                    dst_pitch * (1 << format) / sizeof(uint32_t),
-                    8 * (1 << format), dst_x, dst_y, width, height, color);
+        if (width == 1 && height == 1) {
+            unsigned int i = (dst_x + dst_y * dst_pitch) * (1 << format);
+            switch (format) {

hw/display/sm501.c static void sm501_2d_operation(SM501State *s) · line 875

Quoted 10 lines
hw/display/sm501.c +10 -0
+            case 0:
+                s->local_mem[dst_base + i] = color & 0xff;
+                break;
+            case 1:
+                *(uint16_t *)&s->local_mem[dst_base + i] = color & 0xffff;
+                break;
+            case 2:
+                *(uint32_t *)&s->local_mem[dst_base + i] = color;
+                break;
+            }
              stn_he_p(&s->local_mem[dst_base + i], 1 << format, color);
#14 lore raw
In reply to BALATON Zoltan
Quoted 1 line · BALATON Zoltan

Some small fixes and optimisations for hw/display/sm501.c.

Ping?

Regards,
BALATON Zoltan

Quoted 10 lines

BALATON Zoltan (4):
sm501: Fix bounds checks
sm501: Drop unneded variable
sm501: Ignore no-op blits
sm501: Optimise 1 pixel 2d ops

hw/display/sm501.c | 58 ++++++++++++++++++++++++++++++++++++++--------

1 file changed, 48 insertions(+), 10 deletions(-)