Thread (49 messages) 49 messages, 4 authors, 2021-03-16

Re: [igt-dev] [PATCH i-g-t v23 37/37] tests/gem_linear_blits: Use intel allocator

From: Zbigniew Kempczyński <hidden>
Date: 2021-03-16 13:08:59

On Tue, Mar 16, 2021 at 11:23:25AM +0000, Chris Wilson wrote:
quoted hunk ↗ jump to hunk
Quoting Zbigniew Kempczyński (2021-03-15 16:58:42)
quoted
From: Dominik Grzegorzek <redacted>

Use intel allocator directly, without intel-bb infrastructure.

Signed-off-by: Dominik Grzegorzek <redacted>
Cc: Zbigniew Kempczyński <redacted>
Cc: Chris Wilson <redacted>
---
 tests/i915/gem_linear_blits.c | 120 ++++++++++++++++++++++++----------
 1 file changed, 84 insertions(+), 36 deletions(-)
diff --git a/tests/i915/gem_linear_blits.c b/tests/i915/gem_linear_blits.c
index cae42d52a..b85fd9fef 100644
--- a/tests/i915/gem_linear_blits.c
+++ b/tests/i915/gem_linear_blits.c
@@ -53,10 +53,13 @@ IGT_TEST_DESCRIPTION("Test doing many blits with a working set larger than the"
 #define WIDTH 512
 #define HEIGHT 512
 
+/* We don't have alignment detection yet, so assume worst case scenario */
+#define ALIGNMENT (2048*1024)
+
 static uint32_t linear[WIDTH*HEIGHT];
 
-static void
-copy(int fd, uint32_t dst, uint32_t src)
+static void copy(int fd, uint64_t ahnd, uint32_t dst, uint32_t src,
+                uint64_t dst_offset, uint64_t src_offset, bool do_relocs)
 {
        uint32_t batch[12];
        struct drm_i915_gem_relocation_entry reloc[2];
@@ -64,6 +67,23 @@ copy(int fd, uint32_t dst, uint32_t src)
        struct drm_i915_gem_execbuffer2 exec;
        int i = 0;
 
+       memset(obj, 0, sizeof(obj));
+       obj[0].handle = dst;
+       obj[0].offset = CANONICAL(dst_offset);
+       obj[0].flags = EXEC_OBJECT_PINNED | EXEC_OBJECT_WRITE |
+                      EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
+       obj[1].handle = src;
+       obj[1].offset = CANONICAL(src_offset);
+       obj[1].flags = EXEC_OBJECT_PINNED |
+                      EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
+
+       obj[2].handle = gem_create(fd, 4096);
+       obj[2].offset = intel_allocator_alloc(ahnd, obj[2].handle,
+                       4096, ALIGNMENT);
+       obj[2].offset = CANONICAL(obj[2].offset);
+       obj[2].flags = EXEC_OBJECT_PINNED |
+                      EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
+
        batch[i++] = XY_SRC_COPY_BLT_CMD |
                  XY_SRC_COPY_BLT_WRITE_ALPHA |
                  XY_SRC_COPY_BLT_WRITE_RGB;
@@ -77,49 +97,52 @@ copy(int fd, uint32_t dst, uint32_t src)
                  WIDTH*4;
        batch[i++] = 0; /* dst x1,y1 */
        batch[i++] = (HEIGHT << 16) | WIDTH; /* dst x2,y2 */
-       batch[i++] = 0; /* dst reloc */
+       batch[i++] = obj[0].offset;
        if (intel_gen(intel_get_drm_devid(fd)) >= 8)
-               batch[i++] = 0;
+               batch[i++] = obj[0].offset >> 32;
        batch[i++] = 0; /* src x1,y1 */
        batch[i++] = WIDTH*4;
-       batch[i++] = 0; /* src reloc */
+       batch[i++] = obj[1].offset;
        if (intel_gen(intel_get_drm_devid(fd)) >= 8)
-               batch[i++] = 0;
+               batch[i++] = obj[1].offset >> 32;
        batch[i++] = MI_BATCH_BUFFER_END;
        batch[i++] = MI_NOOP;
 
-       memset(reloc, 0, sizeof(reloc));
-       reloc[0].target_handle = dst;
-       reloc[0].delta = 0;
-       reloc[0].offset = 4 * sizeof(batch[0]);
-       reloc[0].presumed_offset = 0;
-       reloc[0].read_domains = I915_GEM_DOMAIN_RENDER;
-       reloc[0].write_domain = I915_GEM_DOMAIN_RENDER;
-
-       reloc[1].target_handle = src;
-       reloc[1].delta = 0;
-       reloc[1].offset = 7 * sizeof(batch[0]);
-       if (intel_gen(intel_get_drm_devid(fd)) >= 8)
-               reloc[1].offset += sizeof(batch[0]);
-       reloc[1].presumed_offset = 0;
-       reloc[1].read_domains = I915_GEM_DOMAIN_RENDER;
-       reloc[1].write_domain = 0;
-
-       memset(obj, 0, sizeof(obj));
-       obj[0].handle = dst;
-       obj[1].handle = src;
-       obj[2].handle = gem_create(fd, 4096);
        gem_write(fd, obj[2].handle, 0, batch, i * sizeof(batch[0]));
-       obj[2].relocation_count = 2;
-       obj[2].relocs_ptr = to_user_pointer(reloc);
+
+       if (do_relocs) {
+               memset(reloc, 0, sizeof(reloc));
+               reloc[0].target_handle = dst;
+               reloc[0].delta = 0;
+               reloc[0].offset = 4 * sizeof(batch[0]);
+               reloc[0].presumed_offset = obj[0].offset;
+               reloc[0].read_domains = I915_GEM_DOMAIN_RENDER;
+               reloc[0].write_domain = I915_GEM_DOMAIN_RENDER;
+
+               reloc[1].target_handle = src;
+               reloc[1].delta = 0;
+               reloc[1].offset = 7 * sizeof(batch[0]);
+               if (intel_gen(intel_get_drm_devid(fd)) >= 8)
+                       reloc[1].offset += sizeof(batch[0]);
+               reloc[1].presumed_offset = obj[1].offset;
+               reloc[1].read_domains = I915_GEM_DOMAIN_RENDER;
+               reloc[1].write_domain = 0;
+
+               obj[0].flags &= ~EXEC_OBJECT_PINNED;
+               obj[1].flags &= ~EXEC_OBJECT_PINNED;
+               obj[2].flags &= ~EXEC_OBJECT_PINNED;
+               obj[2].relocation_count = ARRAY_SIZE(reloc);
+               obj[2].relocs_ptr = to_user_pointer(reloc);
+       }
 
        memset(&exec, 0, sizeof(exec));
        exec.buffers_ptr = to_user_pointer(obj);
-       exec.buffer_count = 3;
+       exec.buffer_count = ARRAY_SIZE(obj);
        exec.batch_len = i * sizeof(batch[0]);
        exec.flags = gem_has_blt(fd) ? I915_EXEC_BLT : 0;
-
        gem_execbuf(fd, &exec);
+
+       intel_allocator_free(ahnd, obj[2].handle);
        gem_close(fd, obj[2].handle);
 }
@@ -64,6 +67,20 @@ copy(int fd, uint32_t dst, uint32_t src)
 	struct drm_i915_gem_execbuffer2 exec;
 	int i = 0;
 
+	memset(obj, 0, sizeof(obj));
+	obj[0].handle = dst;
+	obj[0].offset = CANONICAL(dst_offset);
+	obj[0].flags = EXEC_OBJECT_WRITE | EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
+	obj[1].handle = src;
+	obj[1].offset = CANONICAL(src_offset);
+	obj[1].flags = EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
+
+	obj[2].handle = gem_create(fd, 4096);
+	obj[2].offset = intel_allocator_alloc(ahnd, obj[2].handle,
+			4096, ALIGNMENT);
+	obj[2].offset = CANONICAL(obj[2].offset);
+	obj[2].flags = EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
+
 	batch[i++] = XY_SRC_COPY_BLT_CMD |
 		  XY_SRC_COPY_BLT_WRITE_ALPHA |
 		  XY_SRC_COPY_BLT_WRITE_RGB;
@@ -77,14 +94,14 @@ copy(int fd, uint32_t dst, uint32_t src)
 		  WIDTH*4;
 	batch[i++] = 0; /* dst x1,y1 */
 	batch[i++] = (HEIGHT << 16) | WIDTH; /* dst x2,y2 */
-	batch[i++] = 0; /* dst reloc */
+	batch[i++] = obj[0].offset;
 	if (intel_gen(intel_get_drm_devid(fd)) >= 8)
-		batch[i++] = 0;
+		batch[i++] = obj[0].offset >> 32;
 	batch[i++] = 0; /* src x1,y1 */
 	batch[i++] = WIDTH*4;
-	batch[i++] = 0; /* src reloc */
+	batch[i++] = obj[1].offset;
 	if (intel_gen(intel_get_drm_devid(fd)) >= 8)
-		batch[i++] = 0;
+		batch[i++] = obj[1].offset >> 32;
 	batch[i++] = MI_BATCH_BUFFER_END;
 	batch[i++] = MI_NOOP;
 
@@ -92,7 +109,7 @@ copy(int fd, uint32_t dst, uint32_t src)
 	reloc[0].target_handle = dst;
 	reloc[0].delta = 0;
 	reloc[0].offset = 4 * sizeof(batch[0]);
-	reloc[0].presumed_offset = 0;
+	reloc[0].presumed_offset = obj[0].offset;
 	reloc[0].read_domains = I915_GEM_DOMAIN_RENDER;
 	reloc[0].write_domain = I915_GEM_DOMAIN_RENDER;
 
@@ -101,25 +118,29 @@ copy(int fd, uint32_t dst, uint32_t src)
 	reloc[1].offset = 7 * sizeof(batch[0]);
 	if (intel_gen(intel_get_drm_devid(fd)) >= 8)
 		reloc[1].offset += sizeof(batch[0]);
-	reloc[1].presumed_offset = 0;
+	reloc[1].presumed_offset = obj[1].offset;
 	reloc[1].read_domains = I915_GEM_DOMAIN_RENDER;
 	reloc[1].write_domain = 0;
 
-	memset(obj, 0, sizeof(obj));
-	obj[0].handle = dst;
-	obj[1].handle = src;
-	obj[2].handle = gem_create(fd, 4096);
 	gem_write(fd, obj[2].handle, 0, batch, i * sizeof(batch[0]));
-	obj[2].relocation_count = 2;
-	obj[2].relocs_ptr = to_user_pointer(reloc);
+
+	if (do_relocs) {
+		obj[2].relocation_count = ARRAY_SIZE(reloc);
+		obj[2].relocs_ptr = to_user_pointer(reloc);
+	} else {
+		obj[0].flags |= EXEC_OBJECT_PINNED;
+		obj[1].flags |= EXEC_OBJECT_PINNED;
+		obj[2].flags |= EXEC_OBJECT_PINNED;
+	}
 
 	memset(&exec, 0, sizeof(exec));
 	exec.buffers_ptr = to_user_pointer(obj);
-	exec.buffer_count = 3;
+	exec.buffer_count = ARRAY_SIZE(obj);
 	exec.batch_len = i * sizeof(batch[0]);
 	exec.flags = gem_has_blt(fd) ? I915_EXEC_BLT : 0;
-
 	gem_execbuf(fd, &exec);
+
+	intel_allocator_free(ahnd, obj[2].handle);
 	gem_close(fd, obj[2].handle);
 }

Is more concise, both in delta and test flow.
Above makes sense, looks better. I'll sent this in an hour.

--
Zbigniew
It also suggests something like

intel_allocator_apply_relocations(ahnd, &exec):
	foreach(exec.obj) {
		foreach(execobj.reloc) {
			if (reloc.write_domain)
				execobj.flags |= EXEC_OBJECT_WRITE;

			if (reloc.presumed_offset == reloc.target->offset)
				continue;

			reloc.presumed_offset = lookup_exec_offset(exec, reloc.target);
			u64 address = reloc.presumed_offset + (int)reloc.delta;
			memcpy(mmap(execobj) + reloc.offset, &address, ahnd->address_size);
		}
		if (!ahnd->use_relocs) {
			execobj.reloc = {};
			execobj.flags |= EXEC_OBJECT_PINNED;
		}
	}
	exec.flags |= I915_EXEC_NORELOC;


would be really useful for converting tests.
I'm going to do this in next series when I start rewriting tests.
Current series has to much iterations and I would like to add 
all convinience code later when this series will be merged.

But yes, you're right and this could save a lot of time during
rewriting.

--
Zbigniew
quoted
@@ -157,17 +180,28 @@ check_bo(int fd, uint32_t handle, uint32_t val)
        igt_assert_eq(num_errors, 0);
 }
 
-static void run_test(int fd, int count)
+static void run_test(int fd, int count, bool do_relocs)
 {
        uint32_t *handle, *start_val;
+       uint64_t *offset, ahnd;
        uint32_t start = 0;
        int i;
 
+       ahnd = intel_allocator_open(fd, 0, do_relocs ?
+                                           INTEL_ALLOCATOR_RELOC :
+                                           INTEL_ALLOCATOR_SIMPLE);
+
        handle = malloc(sizeof(uint32_t) * count * 2);
+       offset = calloc(1, sizeof(uint64_t) * count);
calloc(count, sizeof(uint64_t)
quoted
+       igt_assert_f(handle && offset, "Allocation failed\n");
        start_val = handle + count;
 
        for (i = 0; i < count; i++) {
                handle[i] = create_bo(fd, start);
+
+               offset[i] = intel_allocator_alloc(ahnd, handle[i],
+                                                 sizeof(linear), ALIGNMENT);
+
                start_val[i] = start;
                start += 1024 * 1024 / 4;
        }
@@ -178,17 +212,22 @@ static void run_test(int fd, int count)
 
                if (src == dst)
                        continue;
+               copy(fd, ahnd, handle[dst], handle[src],
+                    offset[dst], offset[src], do_relocs);
 
-               copy(fd, handle[dst], handle[src]);
                start_val[dst] = start_val[src];
        }
 
        for (i = 0; i < count; i++) {
                check_bo(fd, handle[i], start_val[i]);
+               intel_allocator_free(ahnd, handle[i]);
                gem_close(fd, handle[i]);
        }
 
        free(handle);
+       free(offset);
+
+       intel_allocator_close(ahnd);
 }
 
 #define MAX_32b ((1ull << 32) - 4096)
@@ -197,16 +236,21 @@ igt_main
 {
        const int ncpus = sysconf(_SC_NPROCESSORS_ONLN);
        uint64_t count = 0;
+       bool do_relocs;
        int fd = -1;
 
        igt_fixture {
                fd = drm_open_driver(DRIVER_INTEL);
                igt_require_gem(fd);
                gem_require_blitter(fd);
+               do_relocs = !gem_uses_ppgtt(fd);
 
                count = gem_aperture_size(fd);
                if (count >> 32)
                        count = MAX_32b;
+               else
+                       do_relocs = true;
So this catches Braswell (has a 32b full-ppgtt) and only Braswell.
Any reason why the simple allocator cannot handle less than 32b of
address space?
quoted
+
                count = 3 + count / (1024*1024);
                igt_require(count > 1);
                intel_require_memory(count, sizeof(linear), CHECK_RAM);
@@ -216,19 +260,23 @@ igt_main
        }
 
        igt_subtest("basic")
-               run_test(fd, 2);
+               run_test(fd, 2, do_relocs);
 
        igt_subtest("normal") {
+               intel_allocator_multiprocess_start();
                igt_fork(child, ncpus)
-                       run_test(fd, count);
+                       run_test(fd, count, do_relocs);
                igt_waitchildren();
+               intel_allocator_multiprocess_stop();
        }
 
        igt_subtest("interruptible") {
+               intel_allocator_multiprocess_start();
                igt_fork_signal_helper();
                igt_fork(child, ncpus)
-                       run_test(fd, count);
+                       run_test(fd, count, do_relocs);
                igt_waitchildren();
                igt_stop_signal_helper();
+               intel_allocator_multiprocess_stop();
        }
 }
-- 
2.26.0

_______________________________________________
igt-dev mailing list
igt-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/igt-dev
_______________________________________________
igt-dev mailing list
igt-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/igt-dev
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help