};
struct radv_amdgpu_cs {
- struct radeon_winsys_cs base;
+ struct radeon_cmdbuf base;
struct radv_amdgpu_winsys *ws;
struct amdgpu_cs_ib_info ib;
unsigned max_num_buffers;
unsigned num_buffers;
amdgpu_bo_handle *handles;
- uint8_t *priorities;
struct radeon_winsys_bo **old_ib_buffers;
unsigned num_old_ib_buffers;
unsigned num_virtual_buffers;
unsigned max_num_virtual_buffers;
struct radeon_winsys_bo **virtual_buffers;
- uint8_t *virtual_buffer_priorities;
int *virtual_buffer_hash_table;
+
+ /* For chips that don't support chaining. */
+ struct radeon_cmdbuf *old_cs_buffers;
+ unsigned num_old_cs_buffers;
};
static inline struct radv_amdgpu_cs *
-radv_amdgpu_cs(struct radeon_winsys_cs *base)
+radv_amdgpu_cs(struct radeon_cmdbuf *base)
{
return (struct radv_amdgpu_cs*)base;
}
return false;
}
-static void radv_amdgpu_cs_destroy(struct radeon_winsys_cs *rcs)
+static void radv_amdgpu_cs_destroy(struct radeon_cmdbuf *rcs)
{
struct radv_amdgpu_cs *cs = radv_amdgpu_cs(rcs);
for (unsigned i = 0; i < cs->num_old_ib_buffers; ++i)
cs->ws->base.buffer_destroy(cs->old_ib_buffers[i]);
+ for (unsigned i = 0; i < cs->num_old_cs_buffers; ++i) {
+ struct radeon_cmdbuf *rcs = &cs->old_cs_buffers[i];
+ free(rcs->buf);
+ }
+
+ free(cs->old_cs_buffers);
free(cs->old_ib_buffers);
free(cs->virtual_buffers);
- free(cs->virtual_buffer_priorities);
free(cs->virtual_buffer_hash_table);
free(cs->handles);
- free(cs->priorities);
free(cs);
}
cs->hw_ip = ring_to_hw_ip(ring_type);
}
-static struct radeon_winsys_cs *
+static struct radeon_cmdbuf *
radv_amdgpu_cs_create(struct radeon_winsys *ws,
enum ring_type ring_type)
{
cs->ib_size_ptr = &cs->ib.size;
cs->ib.size = 0;
- ws->cs_add_buffer(&cs->base, cs->ib_buffer, 8);
+ ws->cs_add_buffer(&cs->base, cs->ib_buffer);
} else {
cs->base.buf = malloc(16384);
cs->base.max_dw = 4096;
return &cs->base;
}
-static void radv_amdgpu_cs_grow(struct radeon_winsys_cs *_cs, size_t min_size)
+static void radv_amdgpu_cs_grow(struct radeon_cmdbuf *_cs, size_t min_size)
{
struct radv_amdgpu_cs *cs = radv_amdgpu_cs(_cs);
/* The total ib size cannot exceed limit_dws dwords. */
if (ib_dws > limit_dws)
{
- cs->failed = true;
+ /* The maximum size in dwords has been reached,
+ * try to allocate a new one.
+ */
+ cs->old_cs_buffers =
+ realloc(cs->old_cs_buffers,
+ (cs->num_old_cs_buffers + 1) * sizeof(*cs->old_cs_buffers));
+ if (!cs->old_cs_buffers) {
+ cs->failed = true;
+ cs->base.cdw = 0;
+ return;
+ }
+
+ /* Store the current one for submitting it later. */
+ cs->old_cs_buffers[cs->num_old_cs_buffers].cdw = cs->base.cdw;
+ cs->old_cs_buffers[cs->num_old_cs_buffers].max_dw = cs->base.max_dw;
+ cs->old_cs_buffers[cs->num_old_cs_buffers].buf = cs->base.buf;
+ cs->num_old_cs_buffers++;
+
+ /* Reset the cs, it will be re-allocated below. */
cs->base.cdw = 0;
- return;
+ cs->base.buf = NULL;
+
+ /* Re-compute the number of dwords to allocate. */
+ ib_dws = MAX2(cs->base.cdw + min_size,
+ MIN2(cs->base.max_dw * 2, limit_dws));
+ if (ib_dws > limit_dws) {
+ fprintf(stderr, "amdgpu: Too high number of "
+ "dwords to allocate\n");
+ cs->failed = true;
+ return;
+ }
}
uint32_t *new_buf = realloc(cs->base.buf, ib_dws * 4);
ib_size = MIN2(ib_size, 0xfffff);
while (!cs->base.cdw || (cs->base.cdw & 7) != 4)
- cs->base.buf[cs->base.cdw++] = 0xffff1000;
+ radeon_emit(&cs->base, 0xffff1000);
*cs->ib_size_ptr |= cs->base.cdw + 4;
cs->ib_buffer = cs->old_ib_buffers[--cs->num_old_ib_buffers];
}
- cs->ws->base.cs_add_buffer(&cs->base, cs->ib_buffer, 8);
+ cs->ws->base.cs_add_buffer(&cs->base, cs->ib_buffer);
+
+ radeon_emit(&cs->base, PKT3(PKT3_INDIRECT_BUFFER_CIK, 2, 0));
+ radeon_emit(&cs->base, radv_amdgpu_winsys_bo(cs->ib_buffer)->base.va);
+ radeon_emit(&cs->base, radv_amdgpu_winsys_bo(cs->ib_buffer)->base.va >> 32);
+ radeon_emit(&cs->base, S_3F2_CHAIN(1) | S_3F2_VALID(1));
- cs->base.buf[cs->base.cdw++] = PKT3(PKT3_INDIRECT_BUFFER_CIK, 2, 0);
- cs->base.buf[cs->base.cdw++] = radv_amdgpu_winsys_bo(cs->ib_buffer)->base.va;
- cs->base.buf[cs->base.cdw++] = radv_amdgpu_winsys_bo(cs->ib_buffer)->base.va >> 32;
- cs->ib_size_ptr = cs->base.buf + cs->base.cdw;
- cs->base.buf[cs->base.cdw++] = S_3F2_CHAIN(1) | S_3F2_VALID(1);
+ cs->ib_size_ptr = cs->base.buf + cs->base.cdw - 1;
cs->base.buf = (uint32_t *)cs->ib_mapped;
cs->base.cdw = 0;
}
-static bool radv_amdgpu_cs_finalize(struct radeon_winsys_cs *_cs)
+static bool radv_amdgpu_cs_finalize(struct radeon_cmdbuf *_cs)
{
struct radv_amdgpu_cs *cs = radv_amdgpu_cs(_cs);
if (cs->ws->use_ib_bos) {
while (!cs->base.cdw || (cs->base.cdw & 7) != 0)
- cs->base.buf[cs->base.cdw++] = 0xffff1000;
+ radeon_emit(&cs->base, 0xffff1000);
*cs->ib_size_ptr |= cs->base.cdw;
return !cs->failed;
}
-static void radv_amdgpu_cs_reset(struct radeon_winsys_cs *_cs)
+static void radv_amdgpu_cs_reset(struct radeon_cmdbuf *_cs)
{
struct radv_amdgpu_cs *cs = radv_amdgpu_cs(_cs);
cs->base.cdw = 0;
cs->num_virtual_buffers = 0;
if (cs->ws->use_ib_bos) {
- cs->ws->base.cs_add_buffer(&cs->base, cs->ib_buffer, 8);
+ cs->ws->base.cs_add_buffer(&cs->base, cs->ib_buffer);
for (unsigned i = 0; i < cs->num_old_ib_buffers; ++i)
cs->ws->base.buffer_destroy(cs->old_ib_buffers[i]);
cs->ib.ib_mc_address = radv_amdgpu_winsys_bo(cs->ib_buffer)->base.va;
cs->ib_size_ptr = &cs->ib.size;
cs->ib.size = 0;
+ } else {
+ for (unsigned i = 0; i < cs->num_old_cs_buffers; ++i) {
+ struct radeon_cmdbuf *rcs = &cs->old_cs_buffers[i];
+ free(rcs->buf);
+ }
+
+ free(cs->old_cs_buffers);
+ cs->old_cs_buffers = NULL;
+ cs->num_old_cs_buffers = 0;
}
}
}
static void radv_amdgpu_cs_add_buffer_internal(struct radv_amdgpu_cs *cs,
- amdgpu_bo_handle bo,
- uint8_t priority)
+ amdgpu_bo_handle bo)
{
unsigned hash;
int index = radv_amdgpu_cs_find_buffer(cs, bo);
- if (index != -1) {
- cs->priorities[index] = MAX2(cs->priorities[index], priority);
+ if (index != -1)
return;
- }
if (cs->num_buffers == cs->max_num_buffers) {
unsigned new_count = MAX2(1, cs->max_num_buffers * 2);
cs->handles = realloc(cs->handles, new_count * sizeof(amdgpu_bo_handle));
- cs->priorities = realloc(cs->priorities, new_count * sizeof(uint8_t));
cs->max_num_buffers = new_count;
}
cs->handles[cs->num_buffers] = bo;
- cs->priorities[cs->num_buffers] = priority;
hash = ((uintptr_t)bo >> 6) & (ARRAY_SIZE(cs->buffer_hash_table) - 1);
cs->buffer_hash_table[hash] = cs->num_buffers;
++cs->num_buffers;
}
-static void radv_amdgpu_cs_add_virtual_buffer(struct radeon_winsys_cs *_cs,
- struct radeon_winsys_bo *bo,
- uint8_t priority)
+static void radv_amdgpu_cs_add_virtual_buffer(struct radeon_cmdbuf *_cs,
+ struct radeon_winsys_bo *bo)
{
struct radv_amdgpu_cs *cs = radv_amdgpu_cs(_cs);
unsigned hash = ((uintptr_t)bo >> 6) & (VIRTUAL_BUFFER_HASH_TABLE_SIZE - 1);
if (cs->virtual_buffer_hash_table[hash] >= 0) {
int idx = cs->virtual_buffer_hash_table[hash];
if (cs->virtual_buffers[idx] == bo) {
- cs->virtual_buffer_priorities[idx] = MAX2(cs->virtual_buffer_priorities[idx], priority);
return;
}
for (unsigned i = 0; i < cs->num_virtual_buffers; ++i) {
if (cs->virtual_buffers[i] == bo) {
- cs->virtual_buffer_priorities[i] = MAX2(cs->virtual_buffer_priorities[i], priority);
cs->virtual_buffer_hash_table[hash] = i;
return;
}
if(cs->max_num_virtual_buffers <= cs->num_virtual_buffers) {
cs->max_num_virtual_buffers = MAX2(2, cs->max_num_virtual_buffers * 2);
cs->virtual_buffers = realloc(cs->virtual_buffers, sizeof(struct radv_amdgpu_virtual_virtual_buffer*) * cs->max_num_virtual_buffers);
- cs->virtual_buffer_priorities = realloc(cs->virtual_buffer_priorities, sizeof(uint8_t) * cs->max_num_virtual_buffers);
}
cs->virtual_buffers[cs->num_virtual_buffers] = bo;
- cs->virtual_buffer_priorities[cs->num_virtual_buffers] = priority;
cs->virtual_buffer_hash_table[hash] = cs->num_virtual_buffers;
++cs->num_virtual_buffers;
}
-static void radv_amdgpu_cs_add_buffer(struct radeon_winsys_cs *_cs,
- struct radeon_winsys_bo *_bo,
- uint8_t priority)
+static void radv_amdgpu_cs_add_buffer(struct radeon_cmdbuf *_cs,
+ struct radeon_winsys_bo *_bo)
{
struct radv_amdgpu_cs *cs = radv_amdgpu_cs(_cs);
struct radv_amdgpu_winsys_bo *bo = radv_amdgpu_winsys_bo(_bo);
if (bo->is_virtual) {
- radv_amdgpu_cs_add_virtual_buffer(_cs, _bo, priority);
+ radv_amdgpu_cs_add_virtual_buffer(_cs, _bo);
return;
}
if (bo->base.is_local)
return;
- radv_amdgpu_cs_add_buffer_internal(cs, bo->bo, priority);
+ radv_amdgpu_cs_add_buffer_internal(cs, bo->bo);
}
-static void radv_amdgpu_cs_execute_secondary(struct radeon_winsys_cs *_parent,
- struct radeon_winsys_cs *_child)
+static void radv_amdgpu_cs_execute_secondary(struct radeon_cmdbuf *_parent,
+ struct radeon_cmdbuf *_child)
{
struct radv_amdgpu_cs *parent = radv_amdgpu_cs(_parent);
struct radv_amdgpu_cs *child = radv_amdgpu_cs(_child);
for (unsigned i = 0; i < child->num_buffers; ++i) {
- radv_amdgpu_cs_add_buffer_internal(parent, child->handles[i],
- child->priorities[i]);
+ radv_amdgpu_cs_add_buffer_internal(parent, child->handles[i]);
}
for (unsigned i = 0; i < child->num_virtual_buffers; ++i) {
- radv_amdgpu_cs_add_buffer(&parent->base, child->virtual_buffers[i],
- child->virtual_buffer_priorities[i]);
+ radv_amdgpu_cs_add_buffer(&parent->base, child->virtual_buffers[i]);
}
if (parent->ws->use_ib_bos) {
if (parent->base.cdw + 4 > parent->base.max_dw)
radv_amdgpu_cs_grow(&parent->base, 4);
- parent->base.buf[parent->base.cdw++] = PKT3(PKT3_INDIRECT_BUFFER_CIK, 2, 0);
- parent->base.buf[parent->base.cdw++] = child->ib.ib_mc_address;
- parent->base.buf[parent->base.cdw++] = child->ib.ib_mc_address >> 32;
- parent->base.buf[parent->base.cdw++] = child->ib.size;
+ radeon_emit(&parent->base, PKT3(PKT3_INDIRECT_BUFFER_CIK, 2, 0));
+ radeon_emit(&parent->base, child->ib.ib_mc_address);
+ radeon_emit(&parent->base, child->ib.ib_mc_address >> 32);
+ radeon_emit(&parent->base, child->ib.size);
} else {
if (parent->base.cdw + child->base.cdw > parent->base.max_dw)
radv_amdgpu_cs_grow(&parent->base, child->base.cdw);
}
static int radv_amdgpu_create_bo_list(struct radv_amdgpu_winsys *ws,
- struct radeon_winsys_cs **cs_array,
+ struct radeon_cmdbuf **cs_array,
unsigned count,
- struct radv_amdgpu_winsys_bo *extra_bo,
- struct radeon_winsys_cs *extra_cs,
+ struct radv_amdgpu_winsys_bo **extra_bo_array,
+ unsigned num_extra_bo,
+ struct radeon_cmdbuf *extra_cs,
+ const struct radv_winsys_bo_list *radv_bo_list,
amdgpu_bo_list_handle *bo_list)
{
int r = 0;
bo_list);
free(handles);
pthread_mutex_unlock(&ws->global_bo_list_lock);
- } else if (count == 1 && !extra_bo && !extra_cs &&
+ } else if (count == 1 && !num_extra_bo && !extra_cs && !radv_bo_list &&
!radv_amdgpu_cs(cs_array[0])->num_virtual_buffers) {
struct radv_amdgpu_cs *cs = (struct radv_amdgpu_cs*)cs_array[0];
if (cs->num_buffers == 0) {
return 0;
}
r = amdgpu_bo_list_create(ws->dev, cs->num_buffers, cs->handles,
- cs->priorities, bo_list);
+ NULL, bo_list);
} else {
- unsigned total_buffer_count = !!extra_bo;
- unsigned unique_bo_count = !!extra_bo;
+ unsigned total_buffer_count = num_extra_bo;
+ unsigned unique_bo_count = num_extra_bo;
for (unsigned i = 0; i < count; ++i) {
struct radv_amdgpu_cs *cs = (struct radv_amdgpu_cs*)cs_array[i];
total_buffer_count += cs->num_buffers;
if (extra_cs) {
total_buffer_count += ((struct radv_amdgpu_cs*)extra_cs)->num_buffers;
}
+
+ if (radv_bo_list) {
+ total_buffer_count += radv_bo_list->count;
+ }
+
if (total_buffer_count == 0) {
*bo_list = 0;
return 0;
}
amdgpu_bo_handle *handles = malloc(sizeof(amdgpu_bo_handle) * total_buffer_count);
- uint8_t *priorities = malloc(sizeof(uint8_t) * total_buffer_count);
- if (!handles || !priorities) {
+ if (!handles) {
free(handles);
- free(priorities);
return -ENOMEM;
}
- if (extra_bo) {
- handles[0] = extra_bo->bo;
- priorities[0] = 8;
+ for (unsigned i = 0; i < num_extra_bo; i++) {
+ handles[i] = extra_bo_array[i]->bo;
}
for (unsigned i = 0; i < count + !!extra_cs; ++i) {
if (!cs->num_buffers)
continue;
- if (unique_bo_count == 0) {
+ if (unique_bo_count == 0 && !cs->num_virtual_buffers) {
memcpy(handles, cs->handles, cs->num_buffers * sizeof(amdgpu_bo_handle));
- memcpy(priorities, cs->priorities, cs->num_buffers * sizeof(uint8_t));
unique_bo_count = cs->num_buffers;
continue;
}
for (unsigned k = 0; k < unique_bo_so_far; ++k) {
if (handles[k] == cs->handles[j]) {
found = true;
- priorities[k] = MAX2(priorities[k],
- cs->priorities[j]);
break;
}
}
if (!found) {
handles[unique_bo_count] = cs->handles[j];
- priorities[unique_bo_count] = cs->priorities[j];
++unique_bo_count;
}
}
for (unsigned m = 0; m < unique_bo_count; ++m) {
if (handles[m] == bo->bo) {
found = true;
- priorities[m] = MAX2(priorities[m],
- cs->virtual_buffer_priorities[j]);
break;
}
}
if (!found) {
handles[unique_bo_count] = bo->bo;
- priorities[unique_bo_count] = cs->virtual_buffer_priorities[j];
++unique_bo_count;
}
}
}
}
+ if (radv_bo_list) {
+ unsigned unique_bo_so_far = unique_bo_count;
+ for (unsigned i = 0; i < radv_bo_list->count; ++i) {
+ struct radv_amdgpu_winsys_bo *bo = radv_amdgpu_winsys_bo(radv_bo_list->bos[i]);
+ bool found = false;
+ for (unsigned j = 0; j < unique_bo_so_far; ++j) {
+ if (bo->bo == handles[j]) {
+ found = true;
+ break;
+ }
+ }
+ if (!found) {
+ handles[unique_bo_count] = bo->bo;
+ ++unique_bo_count;
+ }
+ }
+ }
+
if (unique_bo_count > 0) {
r = amdgpu_bo_list_create(ws->dev, unique_bo_count, handles,
- priorities, bo_list);
+ NULL, bo_list);
} else {
*bo_list = 0;
}
free(handles);
- free(priorities);
}
return r;
static int radv_amdgpu_winsys_cs_submit_chained(struct radeon_winsys_ctx *_ctx,
int queue_idx,
struct radv_winsys_sem_info *sem_info,
- struct radeon_winsys_cs **cs_array,
+ const struct radv_winsys_bo_list *radv_bo_list,
+ struct radeon_cmdbuf **cs_array,
unsigned cs_count,
- struct radeon_winsys_cs *initial_preamble_cs,
- struct radeon_winsys_cs *continue_preamble_cs,
+ struct radeon_cmdbuf *initial_preamble_cs,
+ struct radeon_cmdbuf *continue_preamble_cs,
struct radeon_winsys_fence *_fence)
{
int r;
amdgpu_bo_list_handle bo_list;
struct amdgpu_cs_request request = {0};
struct amdgpu_cs_ib_info ibs[2];
+ unsigned number_of_ibs = 1;
for (unsigned i = cs_count; i--;) {
struct radv_amdgpu_cs *cs = radv_amdgpu_cs(cs_array[i]);
}
}
- r = radv_amdgpu_create_bo_list(cs0->ws, cs_array, cs_count, NULL, initial_preamble_cs, &bo_list);
+ /* Create a buffer object list. */
+ r = radv_amdgpu_create_bo_list(cs0->ws, cs_array, cs_count, NULL, 0,
+ initial_preamble_cs, radv_bo_list,
+ &bo_list);
if (r) {
fprintf(stderr, "amdgpu: buffer list creation failed for the "
"chained submission(%d)\n", r);
return r;
}
+ /* Configure the CS request. */
+ if (initial_preamble_cs) {
+ ibs[0] = radv_amdgpu_cs(initial_preamble_cs)->ib;
+ ibs[1] = cs0->ib;
+ number_of_ibs++;
+ } else {
+ ibs[0] = cs0->ib;
+ }
+
request.ip_type = cs0->hw_ip;
request.ring = queue_idx;
- request.number_of_ibs = 1;
- request.ibs = &cs0->ib;
+ request.number_of_ibs = number_of_ibs;
+ request.ibs = ibs;
request.resources = bo_list;
request.fence_info = radv_set_cs_fence(ctx, cs0->hw_ip, queue_idx);
- if (initial_preamble_cs) {
- request.ibs = ibs;
- request.number_of_ibs = 2;
- ibs[1] = cs0->ib;
- ibs[0] = ((struct radv_amdgpu_cs*)initial_preamble_cs)->ib;
- }
-
+ /* Submit the CS. */
r = radv_amdgpu_cs_submit(ctx, &request, sem_info);
if (r) {
if (r == -ENOMEM)
if (bo_list)
amdgpu_bo_list_destroy(bo_list);
+ if (r)
+ return r;
+
if (fence)
radv_amdgpu_request_to_fence(ctx, fence, &request);
radv_assign_last_submit(ctx, &request);
- return r;
+ return 0;
}
static int radv_amdgpu_winsys_cs_submit_fallback(struct radeon_winsys_ctx *_ctx,
int queue_idx,
struct radv_winsys_sem_info *sem_info,
- struct radeon_winsys_cs **cs_array,
+ const struct radv_winsys_bo_list *radv_bo_list,
+ struct radeon_cmdbuf **cs_array,
unsigned cs_count,
- struct radeon_winsys_cs *initial_preamble_cs,
- struct radeon_winsys_cs *continue_preamble_cs,
+ struct radeon_cmdbuf *initial_preamble_cs,
+ struct radeon_cmdbuf *continue_preamble_cs,
struct radeon_winsys_fence *_fence)
{
int r;
struct radv_amdgpu_ctx *ctx = radv_amdgpu_ctx(_ctx);
struct radv_amdgpu_fence *fence = (struct radv_amdgpu_fence *)_fence;
amdgpu_bo_list_handle bo_list;
- struct amdgpu_cs_request request;
- bool emit_signal_sem = sem_info->cs_emit_signal;
+ struct amdgpu_cs_request request = {};
+ struct amdgpu_cs_ib_info *ibs;
+ struct radv_amdgpu_cs *cs0;
+ unsigned number_of_ibs;
+
assert(cs_count);
+ cs0 = radv_amdgpu_cs(cs_array[0]);
- for (unsigned i = 0; i < cs_count;) {
- struct radv_amdgpu_cs *cs0 = radv_amdgpu_cs(cs_array[i]);
- struct amdgpu_cs_ib_info ibs[AMDGPU_CS_MAX_IBS_PER_SUBMIT];
- struct radeon_winsys_cs *preamble_cs = i ? continue_preamble_cs : initial_preamble_cs;
- unsigned cnt = MIN2(AMDGPU_CS_MAX_IBS_PER_SUBMIT - !!preamble_cs,
- cs_count - i);
+ /* Compute the number of IBs for this submit. */
+ number_of_ibs = cs_count + !!initial_preamble_cs;
- memset(&request, 0, sizeof(request));
+ /* Create a buffer object list. */
+ r = radv_amdgpu_create_bo_list(cs0->ws, &cs_array[0], cs_count, NULL, 0,
+ initial_preamble_cs, radv_bo_list,
+ &bo_list);
+ if (r) {
+ fprintf(stderr, "amdgpu: buffer list creation failed "
+ "for the fallback submission (%d)\n", r);
+ return r;
+ }
- r = radv_amdgpu_create_bo_list(cs0->ws, &cs_array[i], cnt, NULL,
- preamble_cs, &bo_list);
- if (r) {
- fprintf(stderr, "amdgpu: buffer list creation failed "
- "for the fallback submission (%d)\n", r);
- return r;
- }
+ ibs = malloc(number_of_ibs * sizeof(*ibs));
+ if (!ibs) {
+ if (bo_list)
+ amdgpu_bo_list_destroy(bo_list);
+ return -ENOMEM;
+ }
- request.ip_type = cs0->hw_ip;
- request.ring = queue_idx;
- request.resources = bo_list;
- request.number_of_ibs = cnt + !!preamble_cs;
- request.ibs = ibs;
- request.fence_info = radv_set_cs_fence(ctx, cs0->hw_ip, queue_idx);
+ /* Configure the CS request. */
+ if (initial_preamble_cs)
+ ibs[0] = radv_amdgpu_cs(initial_preamble_cs)->ib;
- if (preamble_cs) {
- ibs[0] = radv_amdgpu_cs(preamble_cs)->ib;
- }
+ for (unsigned i = 0; i < cs_count; i++) {
+ struct radv_amdgpu_cs *cs = radv_amdgpu_cs(cs_array[i]);
- for (unsigned j = 0; j < cnt; ++j) {
- struct radv_amdgpu_cs *cs = radv_amdgpu_cs(cs_array[i + j]);
- ibs[j + !!preamble_cs] = cs->ib;
+ ibs[i + !!initial_preamble_cs] = cs->ib;
- if (cs->is_chained) {
- *cs->ib_size_ptr -= 4;
- cs->is_chained = false;
- }
+ if (cs->is_chained) {
+ *cs->ib_size_ptr -= 4;
+ cs->is_chained = false;
}
+ }
- sem_info->cs_emit_signal = (i == cs_count - cnt) ? emit_signal_sem : false;
- r = radv_amdgpu_cs_submit(ctx, &request, sem_info);
- if (r) {
- if (r == -ENOMEM)
- fprintf(stderr, "amdgpu: Not enough memory for command submission.\n");
- else
- fprintf(stderr, "amdgpu: The CS has been rejected, "
- "see dmesg for more information.\n");
- }
+ request.ip_type = cs0->hw_ip;
+ request.ring = queue_idx;
+ request.resources = bo_list;
+ request.number_of_ibs = number_of_ibs;
+ request.ibs = ibs;
+ request.fence_info = radv_set_cs_fence(ctx, cs0->hw_ip, queue_idx);
- if (bo_list)
- amdgpu_bo_list_destroy(bo_list);
+ /* Submit the CS. */
+ r = radv_amdgpu_cs_submit(ctx, &request, sem_info);
+ if (r) {
+ if (r == -ENOMEM)
+ fprintf(stderr, "amdgpu: Not enough memory for command submission.\n");
+ else
+ fprintf(stderr, "amdgpu: The CS has been rejected, "
+ "see dmesg for more information.\n");
+ }
- if (r)
- return r;
+ if (bo_list)
+ amdgpu_bo_list_destroy(bo_list);
+ free(ibs);
+
+ if (r)
+ return r;
- i += cnt;
- }
if (fence)
radv_amdgpu_request_to_fence(ctx, fence, &request);
static int radv_amdgpu_winsys_cs_submit_sysmem(struct radeon_winsys_ctx *_ctx,
int queue_idx,
struct radv_winsys_sem_info *sem_info,
- struct radeon_winsys_cs **cs_array,
+ const struct radv_winsys_bo_list *radv_bo_list,
+ struct radeon_cmdbuf **cs_array,
unsigned cs_count,
- struct radeon_winsys_cs *initial_preamble_cs,
- struct radeon_winsys_cs *continue_preamble_cs,
+ struct radeon_cmdbuf *initial_preamble_cs,
+ struct radeon_cmdbuf *continue_preamble_cs,
struct radeon_winsys_fence *_fence)
{
int r;
assert(cs_count);
for (unsigned i = 0; i < cs_count;) {
- struct amdgpu_cs_ib_info ib = {0};
- struct radeon_winsys_bo *bo = NULL;
- struct radeon_winsys_cs *preamble_cs = i ? continue_preamble_cs : initial_preamble_cs;
+ struct amdgpu_cs_ib_info *ibs;
+ struct radeon_winsys_bo **bos;
+ struct radeon_cmdbuf *preamble_cs = i ? continue_preamble_cs : initial_preamble_cs;
+ struct radv_amdgpu_cs *cs = radv_amdgpu_cs(cs_array[i]);
+ unsigned number_of_ibs;
uint32_t *ptr;
unsigned cnt = 0;
unsigned size = 0;
unsigned pad_words = 0;
- if (preamble_cs)
- size += preamble_cs->cdw;
- while (i + cnt < cs_count && 0xffff8 - size >= radv_amdgpu_cs(cs_array[i + cnt])->base.cdw) {
- size += radv_amdgpu_cs(cs_array[i + cnt])->base.cdw;
- ++cnt;
- }
+ /* Compute the number of IBs for this submit. */
+ number_of_ibs = cs->num_old_cs_buffers + 1;
+
+ ibs = malloc(number_of_ibs * sizeof(*ibs));
+ if (!ibs)
+ return -ENOMEM;
- while(!size || (size & 7)) {
- size++;
- pad_words++;
+ bos = malloc(number_of_ibs * sizeof(*bos));
+ if (!bos) {
+ free(ibs);
+ return -ENOMEM;
}
- assert(cnt);
- bo = ws->buffer_create(ws, 4 * size, 4096, RADEON_DOMAIN_GTT,
- RADEON_FLAG_CPU_ACCESS |
- RADEON_FLAG_NO_INTERPROCESS_SHARING |
- RADEON_FLAG_READ_ONLY);
- ptr = ws->buffer_map(bo);
+ if (number_of_ibs > 1) {
+ /* Special path when the maximum size in dwords has
+ * been reached because we need to handle more than one
+ * IB per submit.
+ */
+ struct radeon_cmdbuf **new_cs_array;
+ unsigned idx = 0;
- if (preamble_cs) {
- memcpy(ptr, preamble_cs->buf, preamble_cs->cdw * 4);
- ptr += preamble_cs->cdw;
- }
+ new_cs_array = malloc(cs->num_old_cs_buffers *
+ sizeof(*new_cs_array));
+ assert(new_cs_array);
- for (unsigned j = 0; j < cnt; ++j) {
- struct radv_amdgpu_cs *cs = radv_amdgpu_cs(cs_array[i + j]);
- memcpy(ptr, cs->base.buf, 4 * cs->base.cdw);
- ptr += cs->base.cdw;
+ for (unsigned j = 0; j < cs->num_old_cs_buffers; j++)
+ new_cs_array[idx++] = &cs->old_cs_buffers[j];
+ new_cs_array[idx++] = cs_array[i];
- }
+ for (unsigned j = 0; j < number_of_ibs; j++) {
+ struct radeon_cmdbuf *rcs = new_cs_array[j];
+ bool needs_preamble = preamble_cs && j == 0;
+ unsigned size = 0;
- for (unsigned j = 0; j < pad_words; ++j)
- *ptr++ = pad_word;
+ if (needs_preamble)
+ size += preamble_cs->cdw;
+ size += rcs->cdw;
- memset(&request, 0, sizeof(request));
+ assert(size < 0xffff8);
+ while (!size || (size & 7)) {
+ size++;
+ pad_words++;
+ }
+
+ bos[j] = ws->buffer_create(ws, 4 * size, 4096,
+ RADEON_DOMAIN_GTT,
+ RADEON_FLAG_CPU_ACCESS |
+ RADEON_FLAG_NO_INTERPROCESS_SHARING |
+ RADEON_FLAG_READ_ONLY);
+ ptr = ws->buffer_map(bos[j]);
+
+ if (needs_preamble) {
+ memcpy(ptr, preamble_cs->buf, preamble_cs->cdw * 4);
+ ptr += preamble_cs->cdw;
+ }
+
+ memcpy(ptr, rcs->buf, 4 * rcs->cdw);
+ ptr += rcs->cdw;
+
+ for (unsigned k = 0; k < pad_words; ++k)
+ *ptr++ = pad_word;
+
+ ibs[j].size = size;
+ ibs[j].ib_mc_address = radv_buffer_get_va(bos[j]);
+ }
+
+ cnt++;
+ free(new_cs_array);
+ } else {
+ if (preamble_cs)
+ size += preamble_cs->cdw;
+
+ while (i + cnt < cs_count && 0xffff8 - size >= radv_amdgpu_cs(cs_array[i + cnt])->base.cdw) {
+ size += radv_amdgpu_cs(cs_array[i + cnt])->base.cdw;
+ ++cnt;
+ }
+
+ while (!size || (size & 7)) {
+ size++;
+ pad_words++;
+ }
+ assert(cnt);
+
+ bos[0] = ws->buffer_create(ws, 4 * size, 4096,
+ RADEON_DOMAIN_GTT,
+ RADEON_FLAG_CPU_ACCESS |
+ RADEON_FLAG_NO_INTERPROCESS_SHARING |
+ RADEON_FLAG_READ_ONLY);
+ ptr = ws->buffer_map(bos[0]);
+
+ if (preamble_cs) {
+ memcpy(ptr, preamble_cs->buf, preamble_cs->cdw * 4);
+ ptr += preamble_cs->cdw;
+ }
+
+ for (unsigned j = 0; j < cnt; ++j) {
+ struct radv_amdgpu_cs *cs = radv_amdgpu_cs(cs_array[i + j]);
+ memcpy(ptr, cs->base.buf, 4 * cs->base.cdw);
+ ptr += cs->base.cdw;
+
+ }
+
+ for (unsigned j = 0; j < pad_words; ++j)
+ *ptr++ = pad_word;
+
+ ibs[0].size = size;
+ ibs[0].ib_mc_address = radv_buffer_get_va(bos[0]);
+ }
r = radv_amdgpu_create_bo_list(cs0->ws, &cs_array[i], cnt,
- (struct radv_amdgpu_winsys_bo*)bo,
- preamble_cs, &bo_list);
+ (struct radv_amdgpu_winsys_bo **)bos,
+ number_of_ibs, preamble_cs,
+ radv_bo_list, &bo_list);
if (r) {
fprintf(stderr, "amdgpu: buffer list creation failed "
"for the sysmem submission (%d)\n", r);
+ free(ibs);
+ free(bos);
return r;
}
- ib.size = size;
- ib.ib_mc_address = radv_buffer_get_va(bo);
+ memset(&request, 0, sizeof(request));
request.ip_type = cs0->hw_ip;
request.ring = queue_idx;
request.resources = bo_list;
- request.number_of_ibs = 1;
- request.ibs = &ib;
+ request.number_of_ibs = number_of_ibs;
+ request.ibs = ibs;
request.fence_info = radv_set_cs_fence(ctx, cs0->hw_ip, queue_idx);
sem_info->cs_emit_signal = (i == cs_count - cnt) ? emit_signal_sem : false;
if (bo_list)
amdgpu_bo_list_destroy(bo_list);
- ws->buffer_destroy(bo);
+ for (unsigned j = 0; j < number_of_ibs; j++) {
+ ws->buffer_destroy(bos[j]);
+ }
+
+ free(ibs);
+ free(bos);
+
if (r)
return r;
static int radv_amdgpu_winsys_cs_submit(struct radeon_winsys_ctx *_ctx,
int queue_idx,
- struct radeon_winsys_cs **cs_array,
+ struct radeon_cmdbuf **cs_array,
unsigned cs_count,
- struct radeon_winsys_cs *initial_preamble_cs,
- struct radeon_winsys_cs *continue_preamble_cs,
+ struct radeon_cmdbuf *initial_preamble_cs,
+ struct radeon_cmdbuf *continue_preamble_cs,
struct radv_winsys_sem_info *sem_info,
+ const struct radv_winsys_bo_list *bo_list,
bool can_patch,
struct radeon_winsys_fence *_fence)
{
assert(sem_info);
if (!cs->ws->use_ib_bos) {
- ret = radv_amdgpu_winsys_cs_submit_sysmem(_ctx, queue_idx, sem_info, cs_array,
+ ret = radv_amdgpu_winsys_cs_submit_sysmem(_ctx, queue_idx, sem_info, bo_list, cs_array,
cs_count, initial_preamble_cs, continue_preamble_cs, _fence);
- } else if (can_patch && cs_count > AMDGPU_CS_MAX_IBS_PER_SUBMIT && cs->ws->batchchain) {
- ret = radv_amdgpu_winsys_cs_submit_chained(_ctx, queue_idx, sem_info, cs_array,
+ } else if (can_patch && cs->ws->batchchain) {
+ ret = radv_amdgpu_winsys_cs_submit_chained(_ctx, queue_idx, sem_info, bo_list, cs_array,
cs_count, initial_preamble_cs, continue_preamble_cs, _fence);
} else {
- ret = radv_amdgpu_winsys_cs_submit_fallback(_ctx, queue_idx, sem_info, cs_array,
+ ret = radv_amdgpu_winsys_cs_submit_fallback(_ctx, queue_idx, sem_info, bo_list, cs_array,
cs_count, initial_preamble_cs, continue_preamble_cs, _fence);
}
return ret;
}
-static void radv_amdgpu_winsys_cs_dump(struct radeon_winsys_cs *_cs,
+static void radv_amdgpu_winsys_cs_dump(struct radeon_cmdbuf *_cs,
FILE* file,
const int *trace_ids, int trace_id_count)
{
}
if (sem_info->wait.sem_count && sem_info->cs_emit_wait) {
- sem_dependencies = malloc(sizeof(struct drm_amdgpu_cs_chunk_dep) * sem_info->wait.sem_count);
- if (!sem_dependencies) {
- r = -ENOMEM;
- goto error_out;
- }
+ sem_dependencies = alloca(sizeof(struct drm_amdgpu_cs_chunk_dep) * sem_info->wait.sem_count);
int sem_count = 0;
+
for (unsigned j = 0; j < sem_info->wait.sem_count; j++) {
sem = (struct amdgpu_cs_fence *)sem_info->wait.sem[j];
if (!sem->context)
chunks,
&request->seq_no);
error_out:
- free(sem_dependencies);
free(wait_syncobj);
free(signal_syncobj);
return r;