[Mesa-dev] [PATCH 2/2] intel: aubinator: fix decoding with softpin
Lionel Landwerlin
lionel.g.landwerlin at intel.com
Sun Jun 10 12:15:10 UTC 2018
Now that we're softpinning the address of our BOs in anv & i965, the
addresses selected start at the top of the addressing space. This is a
problem for the current implementation of aubinator which uses only a
40bit mmapped address space.
This change keeps track of all the memory writes from the aub file and
fetch them on request by the batch decoder. As a result we can get rid
of the 1<<40 mmapped address space and only rely on the mmap aub file
\o/
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin at intel.com>
---
src/intel/tools/aubinator.c | 116 +++++++++++++++++++++---------------
1 file changed, 68 insertions(+), 48 deletions(-)
diff --git a/src/intel/tools/aubinator.c b/src/intel/tools/aubinator.c
index 3120e82b22e..4a60a606f63 100644
--- a/src/intel/tools/aubinator.c
+++ b/src/intel/tools/aubinator.c
@@ -68,8 +68,14 @@ char *input_file = NULL, *xml_path = NULL;
struct gen_device_info devinfo;
struct gen_batch_decode_ctx batch_ctx;
-uint64_t gtt_size, gtt_end;
-void *gtt;
+static struct memory_chunk {
+ void *map;
+ uint64_t address;
+ uint32_t size;
+} *mem_chunks;
+static uint32_t n_mem_chunks = 0;
+static uint32_t n_allocated_mem_chunks = 0;
+
uint64_t general_state_base;
uint64_t surface_state_base;
uint64_t dynamic_state_base;
@@ -78,6 +84,32 @@ uint64_t instruction_bound;
FILE *outfile;
+static void
+add_mem_chunk(void *data, uint64_t address, uint32_t size)
+{
+ if (n_mem_chunks >= n_allocated_mem_chunks) {
+ n_allocated_mem_chunks = n_allocated_mem_chunks == 0 ? 4 : (2 * n_allocated_mem_chunks);
+ mem_chunks = realloc(mem_chunks, sizeof(*mem_chunks) * n_allocated_mem_chunks);
+ }
+
+ for (int c = 0; c < n_mem_chunks; c++) {
+ struct memory_chunk *chunk = &mem_chunks[c];
+ if (address <= chunk->address && size >= chunk->size) {
+ memmove(chunk, &mem_chunks[c + 1],
+ (n_mem_chunks - c - 1) * sizeof(struct memory_chunk));
+ mem_chunks[n_mem_chunks - 1].map = data;
+ mem_chunks[n_mem_chunks - 1].address = address & ((1ULL << 48) - 1);
+ mem_chunks[n_mem_chunks - 1].size = size;
+ return;
+ }
+ }
+
+ mem_chunks[n_mem_chunks].map = data;
+ mem_chunks[n_mem_chunks].address = address & ((1ULL << 48) - 1);
+ mem_chunks[n_mem_chunks].size = size;
+ n_mem_chunks++;
+}
+
static inline uint32_t
field(uint32_t value, int start, int end)
{
@@ -90,12 +122,6 @@ field(uint32_t value, int start, int end)
struct brw_instruction;
-static inline int
-valid_offset(uint32_t offset)
-{
- return offset < gtt_end;
-}
-
#define GEN_ENGINE_RENDER 1
#define GEN_ENGINE_BLITTER 2
@@ -118,13 +144,7 @@ handle_trace_block(uint32_t *p)
case AUB_TRACE_OP_DATA_WRITE:
if (address_space != AUB_TRACE_MEMTYPE_GTT)
break;
- if (gtt_size < offset + size) {
- fprintf(stderr, "overflow gtt space: %s\n", strerror(errno));
- exit(EXIT_FAILURE);
- }
- memcpy((char *) gtt + offset, data, size);
- if (gtt_end < offset + size)
- gtt_end = offset + size;
+ add_mem_chunk(data, offset, size);
break;
case AUB_TRACE_OP_COMMAND_WRITE:
switch (type) {
@@ -142,7 +162,6 @@ handle_trace_block(uint32_t *p)
(void)engine; /* TODO */
gen_print_batch(&batch_ctx, data, size, 0);
- gtt_end = 0;
break;
}
}
@@ -150,15 +169,19 @@ handle_trace_block(uint32_t *p)
static struct gen_batch_decode_bo
get_gen_batch_bo(void *user_data, uint64_t address)
{
- if (address > gtt_end)
- return (struct gen_batch_decode_bo) { .map = NULL };
-
- /* We really only have one giant address range */
- return (struct gen_batch_decode_bo) {
- .addr = 0,
- .map = gtt,
- .size = gtt_size
- };
+ for (int c = 0; c < n_mem_chunks; c++) {
+ struct memory_chunk *chunk = &mem_chunks[n_mem_chunks - 1 - c];
+ if (chunk->address <= address &&
+ (chunk->address + chunk->size) > address) {
+ return (struct gen_batch_decode_bo) {
+ .addr = chunk->address,
+ .map = chunk->map,
+ .size = chunk->size,
+ };
+ }
+ }
+
+ return (struct gen_batch_decode_bo) { .map = NULL };
}
static void
@@ -252,7 +275,7 @@ handle_memtrace_reg_write(uint32_t *p)
static int blitter_elsp_writes = 0;
static int render_elsq0 = 0;
static int blitter_elsq0 = 0;
- uint8_t *pphwsp;
+ uint32_t pphwsp_offset;
if (offset == 0x2230) {
render_elsp_writes++;
@@ -270,25 +293,38 @@ handle_memtrace_reg_write(uint32_t *p)
return;
}
+
+
if (render_elsp_writes > 3 || blitter_elsp_writes > 3) {
render_elsp_writes = blitter_elsp_writes = 0;
- pphwsp = (uint8_t*)gtt + (value & 0xfffff000);
+ pphwsp_offset = value & 0xfffff000;
} else if (offset == 0x2550) {
engine = GEN_ENGINE_RENDER;
- pphwsp = (uint8_t*)gtt + (render_elsq0 & 0xfffff000);
+ pphwsp_offset = render_elsq0 & 0xfffff000;
} else if (offset == 0x22550) {
engine = GEN_ENGINE_BLITTER;
- pphwsp = (uint8_t*)gtt + (blitter_elsq0 & 0xfffff000);
+ pphwsp_offset = blitter_elsq0 & 0xfffff000;
} else {
return;
}
+ struct gen_batch_decode_bo pphwsp_bo = get_gen_batch_bo(NULL, pphwsp_offset);
+ if (!pphwsp_bo.addr)
+ fprintf(stderr, "Unable to find HW context image\n");
+
+ const void *pphwsp = pphwsp_bo.map + (pphwsp_offset - pphwsp_bo.addr);
const uint32_t pphwsp_size = 4096;
- uint32_t *context = (uint32_t*)(pphwsp + pphwsp_size);
+ const uint32_t *context = (const uint32_t*)(pphwsp + pphwsp_size);
uint32_t ring_buffer_head = context[5];
uint32_t ring_buffer_tail = context[7];
uint32_t ring_buffer_start = context[9];
- uint32_t *commands = (uint32_t*)((uint8_t*)gtt + ring_buffer_start + ring_buffer_head);
+
+ struct gen_batch_decode_bo ring_bo = get_gen_batch_bo(NULL, ring_buffer_start);
+ if (!ring_bo.addr)
+ fprintf(stderr, "Unable to find context ringbuffer\n");
+
+ const void *ring_ptr = ring_bo.map + (ring_buffer_start - ring_bo.addr);
+ const uint32_t *commands = (const uint32_t*)(ring_ptr + ring_buffer_head);
(void)engine; /* TODO */
gen_print_batch(&batch_ctx, commands, ring_buffer_tail - ring_buffer_head, 0);
}
@@ -304,14 +340,7 @@ handle_memtrace_mem_write(uint32_t *p)
if (address_space != 1)
return;
- if (gtt_size < address + size) {
- fprintf(stderr, "overflow gtt space: %s\n", strerror(errno));
- exit(EXIT_FAILURE);
- }
-
- memcpy((char *) gtt + address, data, size);
- if (gtt_end < address + size)
- gtt_end = address + size;
+ add_mem_chunk(data, address, size);
}
struct aub_file {
@@ -637,15 +666,6 @@ int main(int argc, char *argv[])
else
file = aub_file_open(input_file);
- /* mmap a terabyte for our gtt space. */
- gtt_size = 1ull << 40;
- gtt = mmap(NULL, gtt_size, PROT_READ | PROT_WRITE,
- MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE, -1, 0);
- if (gtt == MAP_FAILED) {
- fprintf(stderr, "failed to alloc gtt space: %s\n", strerror(errno));
- exit(EXIT_FAILURE);
- }
-
while (aub_file_more_stuff(file)) {
switch (aub_file_decode_batch(file)) {
case AUB_ITEM_DECODE_OK:
--
2.17.1
More information about the mesa-dev
mailing list