#include "amd/common/sid.h"
-#ifndef AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE
-#define AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE (1 << 3)
-#endif
-
DEBUG_GET_ONCE_BOOL_OPTION(noop, "RADEON_NOOP", false)
/* FENCES */
uint64_t seq_no,
uint64_t *user_fence_cpu_address)
{
- struct amdgpu_fence *rfence = (struct amdgpu_fence*)fence;
+ struct amdgpu_fence *afence = (struct amdgpu_fence*)fence;
- rfence->fence.fence = seq_no;
- rfence->user_fence_cpu_address = user_fence_cpu_address;
- util_queue_fence_signal(&rfence->submitted);
+ afence->fence.fence = seq_no;
+ afence->user_fence_cpu_address = user_fence_cpu_address;
+ util_queue_fence_signal(&afence->submitted);
}
static void amdgpu_fence_signalled(struct pipe_fence_handle *fence)
{
- struct amdgpu_fence *rfence = (struct amdgpu_fence*)fence;
+ struct amdgpu_fence *afence = (struct amdgpu_fence*)fence;
- rfence->signalled = true;
- util_queue_fence_signal(&rfence->submitted);
+ afence->signalled = true;
+ util_queue_fence_signal(&afence->submitted);
}
bool amdgpu_fence_wait(struct pipe_fence_handle *fence, uint64_t timeout,
bool absolute)
{
- struct amdgpu_fence *rfence = (struct amdgpu_fence*)fence;
+ struct amdgpu_fence *afence = (struct amdgpu_fence*)fence;
uint32_t expired;
int64_t abs_timeout;
uint64_t *user_fence_cpu;
int r;
- if (rfence->signalled)
+ if (afence->signalled)
return true;
- /* Handle syncobjs. */
- if (amdgpu_fence_is_syncobj(rfence)) {
- /* Absolute timeouts are only be used by BO fences, which aren't
- * backed by syncobjs.
- */
- assert(!absolute);
+ if (absolute)
+ abs_timeout = timeout;
+ else
+ abs_timeout = os_time_get_absolute_timeout(timeout);
- if (amdgpu_cs_syncobj_wait(rfence->ws->dev, &rfence->syncobj, 1,
- timeout, 0, NULL))
+ /* Handle syncobjs. */
+ if (amdgpu_fence_is_syncobj(afence)) {
+ if (amdgpu_cs_syncobj_wait(afence->ws->dev, &afence->syncobj, 1,
+ abs_timeout, 0, NULL))
return false;
- rfence->signalled = true;
+ afence->signalled = true;
return true;
}
- if (absolute)
- abs_timeout = timeout;
- else
- abs_timeout = os_time_get_absolute_timeout(timeout);
-
/* The fence might not have a number assigned if its IB is being
* submitted in the other thread right now. Wait until the submission
* is done. */
- if (!util_queue_fence_wait_timeout(&rfence->submitted, abs_timeout))
+ if (!util_queue_fence_wait_timeout(&afence->submitted, abs_timeout))
return false;
- user_fence_cpu = rfence->user_fence_cpu_address;
+ user_fence_cpu = afence->user_fence_cpu_address;
if (user_fence_cpu) {
- if (*user_fence_cpu >= rfence->fence.fence) {
- rfence->signalled = true;
+ if (*user_fence_cpu >= afence->fence.fence) {
+ afence->signalled = true;
return true;
}
}
/* Now use the libdrm query. */
- r = amdgpu_cs_query_fence_status(&rfence->fence,
+ r = amdgpu_cs_query_fence_status(&afence->fence,
abs_timeout,
AMDGPU_QUERY_FENCE_TIMEOUT_IS_ABSOLUTE,
&expired);
if (expired) {
/* This variable can only transition from false to true, so it doesn't
* matter if threads race for it. */
- rfence->signalled = true;
+ afence->signalled = true;
return true;
}
return false;
amdgpu_ctx_query_reset_status(struct radeon_winsys_ctx *rwctx)
{
struct amdgpu_ctx *ctx = (struct amdgpu_ctx*)rwctx;
- uint32_t result, hangs;
int r;
+ /* Return a failure due to a GPU hang. */
+ if (ctx->ws->info.drm_minor >= 24) {
+ uint64_t flags;
+
+ r = amdgpu_cs_query_reset_state2(ctx->ctx, &flags);
+ if (r) {
+ fprintf(stderr, "amdgpu: amdgpu_cs_query_reset_state failed. (%i)\n", r);
+ return PIPE_NO_RESET;
+ }
+
+ if (flags & AMDGPU_CTX_QUERY2_FLAGS_RESET) {
+ if (flags & AMDGPU_CTX_QUERY2_FLAGS_GUILTY)
+ return PIPE_GUILTY_CONTEXT_RESET;
+ else
+ return PIPE_INNOCENT_CONTEXT_RESET;
+ }
+ } else {
+ uint32_t result, hangs;
+
+ r = amdgpu_cs_query_reset_state(ctx->ctx, &result, &hangs);
+ if (r) {
+ fprintf(stderr, "amdgpu: amdgpu_cs_query_reset_state failed. (%i)\n", r);
+ return PIPE_NO_RESET;
+ }
+
+ switch (result) {
+ case AMDGPU_CTX_GUILTY_RESET:
+ return PIPE_GUILTY_CONTEXT_RESET;
+ case AMDGPU_CTX_INNOCENT_RESET:
+ return PIPE_INNOCENT_CONTEXT_RESET;
+ case AMDGPU_CTX_UNKNOWN_RESET:
+ return PIPE_UNKNOWN_CONTEXT_RESET;
+ }
+ }
+
/* Return a failure due to a rejected command submission. */
if (ctx->ws->num_total_rejected_cs > ctx->initial_num_total_rejected_cs) {
return ctx->num_rejected_cs ? PIPE_GUILTY_CONTEXT_RESET :
PIPE_INNOCENT_CONTEXT_RESET;
}
-
- /* Return a failure due to a GPU hang. */
- r = amdgpu_cs_query_reset_state(ctx->ctx, &result, &hangs);
- if (r) {
- fprintf(stderr, "amdgpu: amdgpu_cs_query_reset_state failed. (%i)\n", r);
- return PIPE_NO_RESET;
- }
-
- switch (result) {
- case AMDGPU_CTX_GUILTY_RESET:
- return PIPE_GUILTY_CONTEXT_RESET;
- case AMDGPU_CTX_INNOCENT_RESET:
- return PIPE_INNOCENT_CONTEXT_RESET;
- case AMDGPU_CTX_UNKNOWN_RESET:
- return PIPE_UNKNOWN_CONTEXT_RESET;
- case AMDGPU_CTX_NO_RESET:
- default:
- return PIPE_NO_RESET;
- }
+ return PIPE_NO_RESET;
}
/* COMMAND SUBMISSION */
cs->ib[IB_MAIN].ip_type != AMDGPU_HW_IP_VCE &&
cs->ib[IB_MAIN].ip_type != AMDGPU_HW_IP_UVD_ENC &&
cs->ib[IB_MAIN].ip_type != AMDGPU_HW_IP_VCN_DEC &&
- cs->ib[IB_MAIN].ip_type != AMDGPU_HW_IP_VCN_ENC;
+ cs->ib[IB_MAIN].ip_type != AMDGPU_HW_IP_VCN_ENC &&
+ cs->ib[IB_MAIN].ip_type != AMDGPU_HW_IP_VCN_JPEG;
}
static bool amdgpu_cs_has_chaining(struct amdgpu_cs *cs)
{
- return cs->ctx->ws->info.chip_class >= CIK &&
- cs->ring_type == RING_GFX;
+ return cs->ctx->ws->info.chip_class >= GFX7 &&
+ (cs->ring_type == RING_GFX || cs->ring_type == RING_COMPUTE);
}
-static unsigned amdgpu_cs_epilog_dws(enum ring_type ring_type)
+static unsigned amdgpu_cs_epilog_dws(struct amdgpu_cs *cs)
{
- if (ring_type == RING_GFX)
+ if (amdgpu_cs_has_chaining(cs))
return 4; /* for chaining */
return 0;
/* We delay adding the backing buffers until we really have to. However,
* we cannot delay accounting for memory use.
*/
- simple_mtx_lock(&bo->u.sparse.commit_lock);
+ simple_mtx_lock(&bo->lock);
list_for_each_entry(struct amdgpu_sparse_backing, backing, &bo->u.sparse.backing, list) {
if (bo->initial_domain & RADEON_DOMAIN_VRAM)
acs->main.base.used_gart += backing->bo->base.size;
}
- simple_mtx_unlock(&bo->u.sparse.commit_lock);
+ simple_mtx_unlock(&bo->lock);
return idx;
}
return index;
}
-static bool amdgpu_ib_new_buffer(struct amdgpu_winsys *ws, struct amdgpu_ib *ib,
+static bool amdgpu_ib_new_buffer(struct amdgpu_winsys *ws,
+ struct amdgpu_ib *ib,
enum ring_type ring_type)
{
struct pb_buffer *pb;
else
buffer_size = 4 *util_next_power_of_two(4 * ib->max_ib_size);
- buffer_size = MIN2(buffer_size, 4 * 512 * 1024);
+ const unsigned min_size = MAX2(ib->max_check_space_size, 8 * 1024 * 4);
+ const unsigned max_size = 512 * 1024 * 4;
- switch (ib->ib_type) {
- case IB_MAIN:
- buffer_size = MAX2(buffer_size, 8 * 1024 * 4);
- break;
- default:
- unreachable("unhandled IB type");
- }
+ buffer_size = MIN2(buffer_size, max_size);
+ buffer_size = MAX2(buffer_size, min_size); /* min_size is more important */
- pb = ws->base.buffer_create(&ws->base, buffer_size,
- ws->info.gart_page_size,
- RADEON_DOMAIN_GTT,
- RADEON_FLAG_NO_INTERPROCESS_SHARING |
- (ring_type == RING_GFX ||
- ring_type == RING_COMPUTE ||
- ring_type == RING_DMA ?
- RADEON_FLAG_READ_ONLY | RADEON_FLAG_GTT_WC : 0));
+ pb = amdgpu_bo_create(ws, buffer_size,
+ ws->info.gart_page_size,
+ RADEON_DOMAIN_GTT,
+ RADEON_FLAG_NO_INTERPROCESS_SHARING |
+ (ring_type == RING_GFX ||
+ ring_type == RING_COMPUTE ||
+ ring_type == RING_DMA ?
+ RADEON_FLAG_32BIT | RADEON_FLAG_GTT_WC : 0));
if (!pb)
return false;
- mapped = ws->base.buffer_map(pb, NULL, PIPE_TRANSFER_WRITE);
+ mapped = amdgpu_bo_map(pb, NULL, PIPE_TRANSFER_WRITE);
if (!mapped) {
pb_reference(&pb, NULL);
return false;
static unsigned amdgpu_ib_max_submit_dwords(enum ib_type ib_type)
{
+ /* The maximum IB size including all chained IBs. */
switch (ib_type) {
case IB_MAIN:
/* Smaller submits means the GPU gets busy sooner and there is less
* http://www.phoronix.com/scan.php?page=article&item=mesa-111-si&num=1
*/
return 20 * 1024;
+ case IB_PARALLEL_COMPUTE:
+ /* Always chain this IB. */
+ return UINT_MAX;
default:
unreachable("bad ib_type");
}
}
-static bool amdgpu_get_new_ib(struct radeon_winsys *ws, struct amdgpu_cs *cs,
+static bool amdgpu_get_new_ib(struct amdgpu_winsys *ws, struct amdgpu_cs *cs,
enum ib_type ib_type)
{
- struct amdgpu_winsys *aws = (struct amdgpu_winsys*)ws;
/* Small IBs are better than big IBs, because the GPU goes idle quicker
* and there is less waiting for buffers and fences. Proof:
* http://www.phoronix.com/scan.php?page=article&item=mesa-111-si&num=1
*/
struct amdgpu_ib *ib = NULL;
struct drm_amdgpu_cs_chunk_ib *info = &cs->csc->ib[ib_type];
- unsigned ib_size = 0;
+ /* This is the minimum size of a contiguous IB. */
+ unsigned ib_size = 4 * 1024 * 4;
switch (ib_type) {
+ case IB_PARALLEL_COMPUTE:
+ ib = &cs->compute_ib;
+ break;
case IB_MAIN:
ib = &cs->main;
- ib_size = 4 * 1024 * 4;
break;
default:
unreachable("unhandled IB type");
}
+ /* Always allocate at least the size of the biggest cs_check_space call,
+ * because precisely the last call might have requested this size.
+ */
+ ib_size = MAX2(ib_size, ib->max_check_space_size);
+
if (!amdgpu_cs_has_chaining(cs)) {
ib_size = MAX2(ib_size,
4 * MIN2(util_next_power_of_two(ib->max_ib_size),
/* Allocate a new buffer for IBs if the current buffer is all used. */
if (!ib->big_ib_buffer ||
ib->used_ib_space + ib_size > ib->big_ib_buffer->size) {
- if (!amdgpu_ib_new_buffer(aws, ib, cs->ring_type))
+ if (!amdgpu_ib_new_buffer(ws, ib, cs->ring_type))
return false;
}
ib->base.current.buf = (uint32_t*)(ib->ib_mapped + ib->used_ib_space);
ib_size = ib->big_ib_buffer->size - ib->used_ib_space;
- ib->base.current.max_dw = ib_size / 4 - amdgpu_cs_epilog_dws(cs->ring_type);
+ ib->base.current.max_dw = ib_size / 4 - amdgpu_cs_epilog_dws(cs);
+ assert(ib->base.current.max_dw >= ib->max_check_space_size / 4);
+ ib->base.gpu_address = info->va_start;
return true;
}
{
amdgpu_set_ib_size(ib);
ib->used_ib_space += ib->base.current.cdw * 4;
- ib->used_ib_space = align(ib->used_ib_space, ws->info.ib_start_alignment);
+ ib->used_ib_space = align(ib->used_ib_space, ws->info.ib_alignment);
ib->max_ib_size = MAX2(ib->max_ib_size, ib->base.prev_dw + ib->base.current.cdw);
}
cs->ib[IB_MAIN].ip_type = AMDGPU_HW_IP_VCN_ENC;
break;
+ case RING_VCN_JPEG:
+ cs->ib[IB_MAIN].ip_type = AMDGPU_HW_IP_VCN_JPEG;
+ break;
+
case RING_COMPUTE:
case RING_GFX:
cs->ib[IB_MAIN].ip_type = ring_type == RING_GFX ? AMDGPU_HW_IP_GFX :
assert(0);
}
+ cs->ib[IB_PARALLEL_COMPUTE].ip_type = AMDGPU_HW_IP_COMPUTE;
+ cs->ib[IB_PARALLEL_COMPUTE].flags = AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE;
+
memset(cs->buffer_indices_hashlist, -1, sizeof(cs->buffer_indices_hashlist));
cs->last_added_bo = NULL;
return true;
}
+static void cleanup_fence_list(struct amdgpu_fence_list *fences)
+{
+ for (unsigned i = 0; i < fences->num; i++)
+ amdgpu_fence_reference(&fences->list[i], NULL);
+ fences->num = 0;
+}
+
static void amdgpu_cs_context_cleanup(struct amdgpu_cs_context *cs)
{
unsigned i;
p_atomic_dec(&cs->sparse_buffers[i].bo->num_cs_references);
amdgpu_winsys_bo_reference(&cs->sparse_buffers[i].bo, NULL);
}
- for (i = 0; i < cs->num_fence_dependencies; i++)
- amdgpu_fence_reference(&cs->fence_dependencies[i], NULL);
- for (i = 0; i < cs->num_syncobj_to_signal; i++)
- amdgpu_fence_reference(&cs->syncobj_to_signal[i], NULL);
+ cleanup_fence_list(&cs->fence_dependencies);
+ cleanup_fence_list(&cs->syncobj_dependencies);
+ cleanup_fence_list(&cs->syncobj_to_signal);
+ cleanup_fence_list(&cs->compute_fence_dependencies);
+ cleanup_fence_list(&cs->compute_start_fence_dependencies);
cs->num_real_buffers = 0;
cs->num_slab_buffers = 0;
cs->num_sparse_buffers = 0;
- cs->num_fence_dependencies = 0;
- cs->num_syncobj_to_signal = 0;
amdgpu_fence_reference(&cs->fence, NULL);
memset(cs->buffer_indices_hashlist, -1, sizeof(cs->buffer_indices_hashlist));
FREE(cs->real_buffers);
FREE(cs->slab_buffers);
FREE(cs->sparse_buffers);
- FREE(cs->fence_dependencies);
- FREE(cs->syncobj_to_signal);
+ FREE(cs->fence_dependencies.list);
+ FREE(cs->syncobj_dependencies.list);
+ FREE(cs->syncobj_to_signal.list);
+ FREE(cs->compute_fence_dependencies.list);
+ FREE(cs->compute_start_fence_dependencies.list);
}
enum ring_type ring_type,
void (*flush)(void *ctx, unsigned flags,
struct pipe_fence_handle **fence),
- void *flush_ctx)
+ void *flush_ctx,
+ bool stop_exec_on_failure)
{
struct amdgpu_ctx *ctx = (struct amdgpu_ctx*)rwctx;
struct amdgpu_cs *cs;
cs->flush_cs = flush;
cs->flush_data = flush_ctx;
cs->ring_type = ring_type;
+ cs->stop_exec_on_failure = stop_exec_on_failure;
struct amdgpu_cs_fence_info fence_info;
fence_info.handle = cs->ctx->user_fence_bo;
amdgpu_cs_chunk_fence_info_to_data(&fence_info, (void*)&cs->fence_chunk);
cs->main.ib_type = IB_MAIN;
+ cs->compute_ib.ib_type = IB_PARALLEL_COMPUTE;
if (!amdgpu_init_cs_context(ctx->ws, &cs->csc1, ring_type)) {
FREE(cs);
cs->csc = &cs->csc1;
cs->cst = &cs->csc2;
- if (!amdgpu_get_new_ib(&ctx->ws->base, cs, IB_MAIN)) {
+ if (!amdgpu_get_new_ib(ctx->ws, cs, IB_MAIN)) {
amdgpu_destroy_cs_context(&cs->csc2);
amdgpu_destroy_cs_context(&cs->csc1);
FREE(cs);
return &cs->main.base;
}
+static struct radeon_cmdbuf *
+amdgpu_cs_add_parallel_compute_ib(struct radeon_cmdbuf *ib,
+ bool uses_gds_ordered_append)
+{
+ struct amdgpu_cs *cs = (struct amdgpu_cs*)ib;
+ struct amdgpu_winsys *ws = cs->ctx->ws;
+
+ if (cs->ring_type != RING_GFX)
+ return NULL;
+
+ /* only one secondary IB can be added */
+ if (cs->compute_ib.ib_mapped)
+ return NULL;
+
+ /* Allocate the compute IB. */
+ if (!amdgpu_get_new_ib(ws, cs, IB_PARALLEL_COMPUTE))
+ return NULL;
+
+ if (uses_gds_ordered_append) {
+ cs->csc1.ib[IB_PARALLEL_COMPUTE].flags |=
+ AMDGPU_IB_FLAG_RESET_GDS_MAX_WAVE_ID;
+ cs->csc2.ib[IB_PARALLEL_COMPUTE].flags |=
+ AMDGPU_IB_FLAG_RESET_GDS_MAX_WAVE_ID;
+ }
+ return &cs->compute_ib.base;
+}
+
+static bool
+amdgpu_cs_setup_preemption(struct radeon_cmdbuf *rcs, const uint32_t *preamble_ib,
+ unsigned preamble_num_dw)
+{
+ struct amdgpu_ib *ib = amdgpu_ib(rcs);
+ struct amdgpu_cs *cs = amdgpu_cs_from_ib(ib);
+ struct amdgpu_winsys *ws = cs->ctx->ws;
+ struct amdgpu_cs_context *csc[2] = {&cs->csc1, &cs->csc2};
+ unsigned size = align(preamble_num_dw * 4, ws->info.ib_alignment);
+ struct pb_buffer *preamble_bo;
+ uint32_t *map;
+
+ /* Create the preamble IB buffer. */
+ preamble_bo = amdgpu_bo_create(ws, size, ws->info.ib_alignment,
+ RADEON_DOMAIN_VRAM,
+ RADEON_FLAG_NO_INTERPROCESS_SHARING |
+ RADEON_FLAG_GTT_WC |
+ RADEON_FLAG_READ_ONLY);
+ if (!preamble_bo)
+ return false;
+
+ map = (uint32_t*)amdgpu_bo_map(preamble_bo, NULL,
+ PIPE_TRANSFER_WRITE | RADEON_TRANSFER_TEMPORARY);
+ if (!map) {
+ pb_reference(&preamble_bo, NULL);
+ return false;
+ }
+
+ /* Upload the preamble IB. */
+ memcpy(map, preamble_ib, preamble_num_dw * 4);
+
+ /* Pad the IB. */
+ uint32_t ib_pad_dw_mask = ws->info.ib_pad_dw_mask[cs->ring_type];
+ while (preamble_num_dw & ib_pad_dw_mask)
+ map[preamble_num_dw++] = PKT3_NOP_PAD;
+ amdgpu_bo_unmap(preamble_bo);
+
+ for (unsigned i = 0; i < 2; i++) {
+ csc[i]->ib[IB_PREAMBLE] = csc[i]->ib[IB_MAIN];
+ csc[i]->ib[IB_PREAMBLE].flags |= AMDGPU_IB_FLAG_PREAMBLE;
+ csc[i]->ib[IB_PREAMBLE].va_start = amdgpu_winsys_bo(preamble_bo)->va;
+ csc[i]->ib[IB_PREAMBLE].ib_bytes = preamble_num_dw * 4;
+
+ csc[i]->ib[IB_MAIN].flags |= AMDGPU_IB_FLAG_PREEMPT;
+ }
+
+ assert(!cs->preamble_ib_bo);
+ cs->preamble_ib_bo = preamble_bo;
+
+ amdgpu_cs_add_buffer(rcs, cs->preamble_ib_bo, RADEON_USAGE_READ, 0,
+ RADEON_PRIO_IB1);
+ return true;
+}
+
static bool amdgpu_cs_validate(struct radeon_cmdbuf *rcs)
{
return true;
}
-static bool amdgpu_cs_check_space(struct radeon_cmdbuf *rcs, unsigned dw)
+static bool amdgpu_cs_check_space(struct radeon_cmdbuf *rcs, unsigned dw,
+ bool force_chaining)
{
struct amdgpu_ib *ib = amdgpu_ib(rcs);
struct amdgpu_cs *cs = amdgpu_cs_from_ib(ib);
unsigned requested_size = rcs->prev_dw + rcs->current.cdw + dw;
+ unsigned cs_epilog_dw = amdgpu_cs_epilog_dws(cs);
+ unsigned need_byte_size = (dw + cs_epilog_dw) * 4;
uint64_t va;
uint32_t *new_ptr_ib_size;
assert(rcs->current.cdw <= rcs->current.max_dw);
- if (requested_size > amdgpu_ib_max_submit_dwords(ib->ib_type))
- return false;
+ /* 125% of the size for IB epilog. */
+ unsigned safe_byte_size = need_byte_size + need_byte_size / 4;
+ ib->max_check_space_size = MAX2(ib->max_check_space_size,
+ safe_byte_size);
- ib->max_ib_size = MAX2(ib->max_ib_size, requested_size);
+ /* If force_chaining is true, we can't return. We have to chain. */
+ if (!force_chaining) {
+ if (requested_size > amdgpu_ib_max_submit_dwords(ib->ib_type))
+ return false;
- if (rcs->current.max_dw - rcs->current.cdw >= dw)
- return true;
+ ib->max_ib_size = MAX2(ib->max_ib_size, requested_size);
- if (!amdgpu_cs_has_chaining(cs))
+ if (rcs->current.max_dw - rcs->current.cdw >= dw)
+ return true;
+ }
+
+ if (!amdgpu_cs_has_chaining(cs)) {
+ assert(!force_chaining);
return false;
+ }
/* Allocate a new chunk */
if (rcs->num_prev >= rcs->max_prev) {
va = amdgpu_winsys_bo(ib->big_ib_buffer)->va;
/* This space was originally reserved. */
- rcs->current.max_dw += 4;
- assert(ib->used_ib_space + 4 * rcs->current.max_dw <= ib->big_ib_buffer->size);
+ rcs->current.max_dw += cs_epilog_dw;
- /* Pad with NOPs and add INDIRECT_BUFFER packet */
- while ((rcs->current.cdw & 7) != 4)
- radeon_emit(rcs, 0xffff1000); /* type3 nop packet */
+ /* Pad with NOPs but leave 4 dwords for INDIRECT_BUFFER. */
+ uint32_t ib_pad_dw_mask = cs->ctx->ws->info.ib_pad_dw_mask[cs->ring_type];
+ while ((rcs->current.cdw & ib_pad_dw_mask) != ib_pad_dw_mask - 3)
+ radeon_emit(rcs, PKT3_NOP_PAD);
- radeon_emit(rcs, PKT3(ib->ib_type == IB_MAIN ? PKT3_INDIRECT_BUFFER_CIK
- : PKT3_INDIRECT_BUFFER_CONST, 2, 0));
+ radeon_emit(rcs, PKT3(PKT3_INDIRECT_BUFFER_CIK, 2, 0));
radeon_emit(rcs, va);
radeon_emit(rcs, va >> 32);
new_ptr_ib_size = &rcs->current.buf[rcs->current.cdw++];
+ assert((rcs->current.cdw & ib_pad_dw_mask) == 0);
assert((rcs->current.cdw & 7) == 0);
assert(rcs->current.cdw <= rcs->current.max_dw);
ib->base.current.cdw = 0;
ib->base.current.buf = (uint32_t*)(ib->ib_mapped + ib->used_ib_space);
- ib->base.current.max_dw = ib->big_ib_buffer->size / 4 - amdgpu_cs_epilog_dws(cs->ring_type);
+ ib->base.current.max_dw = ib->big_ib_buffer->size / 4 - cs_epilog_dw;
+ assert(ib->base.current.max_dw >= ib->max_check_space_size / 4);
+ ib->base.gpu_address = va;
amdgpu_cs_add_buffer(&cs->main.base, ib->big_ib_buffer,
RADEON_USAGE_READ, 0, RADEON_PRIO_IB1);
return cs->num_real_buffers;
}
-static unsigned add_fence_dependency_entry(struct amdgpu_cs_context *cs)
+static void add_fence_to_list(struct amdgpu_fence_list *fences,
+ struct amdgpu_fence *fence)
{
- unsigned idx = cs->num_fence_dependencies++;
+ unsigned idx = fences->num++;
- if (idx >= cs->max_fence_dependencies) {
+ if (idx >= fences->max) {
unsigned size;
const unsigned increment = 8;
- cs->max_fence_dependencies = idx + increment;
- size = cs->max_fence_dependencies * sizeof(cs->fence_dependencies[0]);
- cs->fence_dependencies = realloc(cs->fence_dependencies, size);
+ fences->max = idx + increment;
+ size = fences->max * sizeof(fences->list[0]);
+ fences->list = realloc(fences->list, size);
/* Clear the newly-allocated elements. */
- memset(cs->fence_dependencies + idx, 0,
- increment * sizeof(cs->fence_dependencies[0]));
+ memset(fences->list + idx, 0,
+ increment * sizeof(fences->list[0]));
}
- return idx;
+ amdgpu_fence_reference(&fences->list[idx], (struct pipe_fence_handle*)fence);
}
static bool is_noop_fence_dependency(struct amdgpu_cs *acs,
{
struct amdgpu_cs_context *cs = acs->csc;
- if (!amdgpu_fence_is_syncobj(fence) &&
+ /* Detect no-op dependencies only when there is only 1 ring,
+ * because IBs on one ring are always executed one at a time.
+ *
+ * We always want no dependency between back-to-back gfx IBs, because
+ * we need the parallelism between IBs for good performance.
+ */
+ if ((acs->ring_type == RING_GFX ||
+ acs->ctx->ws->info.num_rings[acs->ring_type] == 1) &&
+ !amdgpu_fence_is_syncobj(fence) &&
fence->ctx == acs->ctx &&
fence->fence.ip_type == cs->ib[IB_MAIN].ip_type &&
fence->fence.ip_instance == cs->ib[IB_MAIN].ip_instance &&
}
static void amdgpu_cs_add_fence_dependency(struct radeon_cmdbuf *rws,
- struct pipe_fence_handle *pfence)
+ struct pipe_fence_handle *pfence,
+ unsigned dependency_flags)
{
struct amdgpu_cs *acs = amdgpu_cs(rws);
struct amdgpu_cs_context *cs = acs->csc;
util_queue_fence_wait(&fence->submitted);
+ if (dependency_flags & RADEON_DEPENDENCY_PARALLEL_COMPUTE_ONLY) {
+ /* Syncobjs are not needed here. */
+ assert(!amdgpu_fence_is_syncobj(fence));
+
+ if (acs->ctx->ws->info.has_scheduled_fence_dependency &&
+ dependency_flags & RADEON_DEPENDENCY_START_FENCE)
+ add_fence_to_list(&cs->compute_start_fence_dependencies, fence);
+ else
+ add_fence_to_list(&cs->compute_fence_dependencies, fence);
+ return;
+ }
+
+ /* Start fences are not needed here. */
+ assert(!(dependency_flags & RADEON_DEPENDENCY_START_FENCE));
+
if (is_noop_fence_dependency(acs, fence))
return;
- unsigned idx = add_fence_dependency_entry(cs);
- amdgpu_fence_reference(&cs->fence_dependencies[idx],
- (struct pipe_fence_handle*)fence);
+ if (amdgpu_fence_is_syncobj(fence))
+ add_fence_to_list(&cs->syncobj_dependencies, fence);
+ else
+ add_fence_to_list(&cs->fence_dependencies, fence);
}
static void amdgpu_add_bo_fence_dependencies(struct amdgpu_cs *acs,
if (!(buffer->usage & RADEON_USAGE_SYNCHRONIZED))
continue;
- unsigned idx = add_fence_dependency_entry(cs);
- amdgpu_fence_reference(&cs->fence_dependencies[idx],
- (struct pipe_fence_handle*)bo_fence);
+ add_fence_to_list(&cs->fence_dependencies, bo_fence);
}
for (unsigned j = new_num_fences; j < bo->num_fences; ++j)
{
struct amdgpu_cs_context *cs = acs->csc;
- cs->num_fence_dependencies = 0;
-
amdgpu_add_fence_dependencies_bo_list(acs, cs->fence, cs->num_real_buffers, cs->real_buffers);
amdgpu_add_fence_dependencies_bo_list(acs, cs->fence, cs->num_slab_buffers, cs->slab_buffers);
amdgpu_add_fence_dependencies_bo_list(acs, cs->fence, cs->num_sparse_buffers, cs->sparse_buffers);
}
-static unsigned add_syncobj_to_signal_entry(struct amdgpu_cs_context *cs)
-{
- unsigned idx = cs->num_syncobj_to_signal++;
-
- if (idx >= cs->max_syncobj_to_signal) {
- unsigned size;
- const unsigned increment = 8;
-
- cs->max_syncobj_to_signal = idx + increment;
- size = cs->max_syncobj_to_signal * sizeof(cs->syncobj_to_signal[0]);
- cs->syncobj_to_signal = realloc(cs->syncobj_to_signal, size);
- /* Clear the newly-allocated elements. */
- memset(cs->syncobj_to_signal + idx, 0,
- increment * sizeof(cs->syncobj_to_signal[0]));
- }
- return idx;
-}
-
static void amdgpu_cs_add_syncobj_signal(struct radeon_cmdbuf *rws,
struct pipe_fence_handle *fence)
{
assert(amdgpu_fence_is_syncobj((struct amdgpu_fence *)fence));
- unsigned idx = add_syncobj_to_signal_entry(cs);
- amdgpu_fence_reference(&cs->syncobj_to_signal[idx], fence);
+ add_fence_to_list(&cs->syncobj_to_signal, (struct amdgpu_fence*)fence);
}
/* Add backing of sparse buffers to the buffer list.
struct amdgpu_cs_buffer *buffer = &cs->sparse_buffers[i];
struct amdgpu_winsys_bo *bo = buffer->bo;
- simple_mtx_lock(&bo->u.sparse.commit_lock);
+ simple_mtx_lock(&bo->lock);
list_for_each_entry(struct amdgpu_sparse_backing, backing, &bo->u.sparse.backing, list) {
/* We can directly add the buffer here, because we know that each
int idx = amdgpu_do_add_real_buffer(cs, backing->bo);
if (idx < 0) {
fprintf(stderr, "%s: failed to add buffer\n", __FUNCTION__);
- simple_mtx_unlock(&bo->u.sparse.commit_lock);
+ simple_mtx_unlock(&bo->lock);
return false;
}
p_atomic_inc(&backing->bo->num_active_ioctls);
}
- simple_mtx_unlock(&bo->u.sparse.commit_lock);
+ simple_mtx_unlock(&bo->lock);
}
return true;
struct amdgpu_winsys *ws = acs->ctx->ws;
struct amdgpu_cs_context *cs = acs->cst;
int i, r;
- amdgpu_bo_list_handle bo_list = NULL;
+ uint32_t bo_list = 0;
uint64_t seq_no = 0;
bool has_user_fence = amdgpu_cs_has_user_fence(cs);
+ bool use_bo_list_create = ws->info.drm_minor < 27;
+ struct drm_amdgpu_bo_list_in bo_list_in;
- /* Create the buffer list.
- * Use a buffer list containing all allocated buffers if requested.
- */
+ /* Prepare the buffer list. */
if (ws->debug_all_bos) {
+ /* The buffer list contains all buffers. This is a slow path that
+ * ensures that no buffer is missing in the BO list.
+ */
+ unsigned num_handles = 0;
+ struct drm_amdgpu_bo_list_entry *list =
+ alloca(ws->num_buffers * sizeof(struct drm_amdgpu_bo_list_entry));
struct amdgpu_winsys_bo *bo;
- amdgpu_bo_handle *handles;
- unsigned num = 0;
simple_mtx_lock(&ws->global_bo_list_lock);
- handles = alloca(sizeof(handles[0]) * ws->num_buffers);
-
LIST_FOR_EACH_ENTRY(bo, &ws->global_bo_list, u.real.global_list_item) {
- assert(num < ws->num_buffers);
- handles[num++] = bo->bo;
+ list[num_handles].bo_handle = bo->u.real.kms_handle;
+ list[num_handles].bo_priority = 0;
+ ++num_handles;
}
- r = amdgpu_bo_list_create(ws->dev, ws->num_buffers,
- handles, NULL, &bo_list);
+ r = amdgpu_bo_list_create_raw(ws->dev, ws->num_buffers, list, &bo_list);
simple_mtx_unlock(&ws->global_bo_list_lock);
if (r) {
fprintf(stderr, "amdgpu: buffer list creation failed (%d)\n", r);
- amdgpu_fence_signalled(cs->fence);
- cs->error_code = r;
goto cleanup;
}
} else {
- unsigned num_handles;
-
if (!amdgpu_add_sparse_backing_buffers(cs)) {
fprintf(stderr, "amdgpu: amdgpu_add_sparse_backing_buffers failed\n");
- amdgpu_fence_signalled(cs->fence);
- cs->error_code = -ENOMEM;
+ r = -ENOMEM;
goto cleanup;
}
- amdgpu_bo_handle *handles = alloca(sizeof(*handles) * cs->num_real_buffers);
- uint8_t *flags = alloca(sizeof(*flags) * cs->num_real_buffers);
+ struct drm_amdgpu_bo_list_entry *list =
+ alloca((cs->num_real_buffers + 2) * sizeof(struct drm_amdgpu_bo_list_entry));
- num_handles = 0;
+ unsigned num_handles = 0;
for (i = 0; i < cs->num_real_buffers; ++i) {
struct amdgpu_cs_buffer *buffer = &cs->real_buffers[i];
-
- if (buffer->bo->is_local)
- continue;
-
assert(buffer->u.real.priority_usage != 0);
- handles[num_handles] = buffer->bo->bo;
- flags[num_handles] = (util_last_bit(buffer->u.real.priority_usage) - 1) / 2;
- ++num_handles;
+ list[num_handles].bo_handle = buffer->bo->u.real.kms_handle;
+ list[num_handles].bo_priority = (util_last_bit(buffer->u.real.priority_usage) - 1) / 2;
+ ++num_handles;
}
- if (num_handles) {
- r = amdgpu_bo_list_create(ws->dev, num_handles,
- handles, flags, &bo_list);
+ if (use_bo_list_create) {
+ /* Legacy path creating the buffer list handle and passing it to the CS ioctl. */
+ r = amdgpu_bo_list_create_raw(ws->dev, num_handles, list, &bo_list);
if (r) {
fprintf(stderr, "amdgpu: buffer list creation failed (%d)\n", r);
- amdgpu_fence_signalled(cs->fence);
- cs->error_code = r;
goto cleanup;
}
+ } else {
+ /* Standard path passing the buffer list via the CS ioctl. */
+ bo_list_in.operation = ~0;
+ bo_list_in.list_handle = ~0;
+ bo_list_in.bo_number = num_handles;
+ bo_list_in.bo_info_size = sizeof(struct drm_amdgpu_bo_list_entry);
+ bo_list_in.bo_info_ptr = (uint64_t)(uintptr_t)list;
}
}
if (acs->ring_type == RING_GFX)
ws->gfx_bo_list_counter += cs->num_real_buffers;
- if (acs->ctx->num_rejected_cs) {
+ if (acs->stop_exec_on_failure && acs->ctx->num_rejected_cs) {
r = -ECANCELED;
} else {
- struct drm_amdgpu_cs_chunk chunks[5];
+ struct drm_amdgpu_cs_chunk chunks[7];
unsigned num_chunks = 0;
- /* Convert from dwords to bytes. */
- cs->ib[IB_MAIN].ib_bytes *= 4;
-
- /* IB */
- chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_IB;
- chunks[num_chunks].length_dw = sizeof(struct drm_amdgpu_cs_chunk_ib) / 4;
- chunks[num_chunks].chunk_data = (uintptr_t)&cs->ib[IB_MAIN];
- num_chunks++;
-
- /* Fence */
- if (has_user_fence) {
- chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_FENCE;
- chunks[num_chunks].length_dw = sizeof(struct drm_amdgpu_cs_chunk_fence) / 4;
- chunks[num_chunks].chunk_data = (uintptr_t)&acs->fence_chunk;
+ /* BO list */
+ if (!use_bo_list_create) {
+ chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_BO_HANDLES;
+ chunks[num_chunks].length_dw = sizeof(struct drm_amdgpu_bo_list_in) / 4;
+ chunks[num_chunks].chunk_data = (uintptr_t)&bo_list_in;
num_chunks++;
}
- /* Dependencies */
- unsigned num_dependencies = cs->num_fence_dependencies;
- unsigned num_syncobj_dependencies = 0;
-
+ /* Fence dependencies. */
+ unsigned num_dependencies = cs->fence_dependencies.num;
if (num_dependencies) {
struct drm_amdgpu_cs_chunk_dep *dep_chunk =
alloca(num_dependencies * sizeof(*dep_chunk));
- unsigned num = 0;
for (unsigned i = 0; i < num_dependencies; i++) {
struct amdgpu_fence *fence =
- (struct amdgpu_fence*)cs->fence_dependencies[i];
-
- if (amdgpu_fence_is_syncobj(fence)) {
- num_syncobj_dependencies++;
- continue;
- }
+ (struct amdgpu_fence*)cs->fence_dependencies.list[i];
assert(util_queue_fence_is_signalled(&fence->submitted));
- amdgpu_cs_chunk_fence_to_dep(&fence->fence, &dep_chunk[num++]);
+ amdgpu_cs_chunk_fence_to_dep(&fence->fence, &dep_chunk[i]);
}
chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_DEPENDENCIES;
- chunks[num_chunks].length_dw = sizeof(dep_chunk[0]) / 4 * num;
+ chunks[num_chunks].length_dw = sizeof(dep_chunk[0]) / 4 * num_dependencies;
chunks[num_chunks].chunk_data = (uintptr_t)dep_chunk;
num_chunks++;
}
/* Syncobj dependencies. */
+ unsigned num_syncobj_dependencies = cs->syncobj_dependencies.num;
if (num_syncobj_dependencies) {
struct drm_amdgpu_cs_chunk_sem *sem_chunk =
alloca(num_syncobj_dependencies * sizeof(sem_chunk[0]));
- unsigned num = 0;
- for (unsigned i = 0; i < num_dependencies; i++) {
+ for (unsigned i = 0; i < num_syncobj_dependencies; i++) {
struct amdgpu_fence *fence =
- (struct amdgpu_fence*)cs->fence_dependencies[i];
+ (struct amdgpu_fence*)cs->syncobj_dependencies.list[i];
if (!amdgpu_fence_is_syncobj(fence))
continue;
assert(util_queue_fence_is_signalled(&fence->submitted));
- sem_chunk[num++].handle = fence->syncobj;
+ sem_chunk[i].handle = fence->syncobj;
}
chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_SYNCOBJ_IN;
- chunks[num_chunks].length_dw = sizeof(sem_chunk[0]) / 4 * num;
+ chunks[num_chunks].length_dw = sizeof(sem_chunk[0]) / 4 * num_syncobj_dependencies;
chunks[num_chunks].chunk_data = (uintptr_t)sem_chunk;
num_chunks++;
}
- /* Syncobj sygnals. */
- if (cs->num_syncobj_to_signal) {
+ /* Submit the parallel compute IB first. */
+ if (cs->ib[IB_PARALLEL_COMPUTE].ib_bytes > 0) {
+ unsigned old_num_chunks = num_chunks;
+
+ /* Add compute fence dependencies. */
+ unsigned num_dependencies = cs->compute_fence_dependencies.num;
+ if (num_dependencies) {
+ struct drm_amdgpu_cs_chunk_dep *dep_chunk =
+ alloca(num_dependencies * sizeof(*dep_chunk));
+
+ for (unsigned i = 0; i < num_dependencies; i++) {
+ struct amdgpu_fence *fence =
+ (struct amdgpu_fence*)cs->compute_fence_dependencies.list[i];
+
+ assert(util_queue_fence_is_signalled(&fence->submitted));
+ amdgpu_cs_chunk_fence_to_dep(&fence->fence, &dep_chunk[i]);
+ }
+
+ chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_DEPENDENCIES;
+ chunks[num_chunks].length_dw = sizeof(dep_chunk[0]) / 4 * num_dependencies;
+ chunks[num_chunks].chunk_data = (uintptr_t)dep_chunk;
+ num_chunks++;
+ }
+
+ /* Add compute start fence dependencies. */
+ unsigned num_start_dependencies = cs->compute_start_fence_dependencies.num;
+ if (num_start_dependencies) {
+ struct drm_amdgpu_cs_chunk_dep *dep_chunk =
+ alloca(num_start_dependencies * sizeof(*dep_chunk));
+
+ for (unsigned i = 0; i < num_start_dependencies; i++) {
+ struct amdgpu_fence *fence =
+ (struct amdgpu_fence*)cs->compute_start_fence_dependencies.list[i];
+
+ assert(util_queue_fence_is_signalled(&fence->submitted));
+ amdgpu_cs_chunk_fence_to_dep(&fence->fence, &dep_chunk[i]);
+ }
+
+ chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES;
+ chunks[num_chunks].length_dw = sizeof(dep_chunk[0]) / 4 * num_start_dependencies;
+ chunks[num_chunks].chunk_data = (uintptr_t)dep_chunk;
+ num_chunks++;
+ }
+
+ /* Convert from dwords to bytes. */
+ cs->ib[IB_PARALLEL_COMPUTE].ib_bytes *= 4;
+ chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_IB;
+ chunks[num_chunks].length_dw = sizeof(struct drm_amdgpu_cs_chunk_ib) / 4;
+ chunks[num_chunks].chunk_data = (uintptr_t)&cs->ib[IB_PARALLEL_COMPUTE];
+ num_chunks++;
+
+ r = amdgpu_cs_submit_raw2(ws->dev, acs->ctx->ctx, bo_list,
+ num_chunks, chunks, NULL);
+ if (r)
+ goto finalize;
+
+ /* Back off the compute chunks. */
+ num_chunks = old_num_chunks;
+ }
+
+ /* Syncobj signals. */
+ unsigned num_syncobj_to_signal = cs->syncobj_to_signal.num;
+ if (num_syncobj_to_signal) {
struct drm_amdgpu_cs_chunk_sem *sem_chunk =
- alloca(cs->num_syncobj_to_signal * sizeof(sem_chunk[0]));
+ alloca(num_syncobj_to_signal * sizeof(sem_chunk[0]));
- for (unsigned i = 0; i < cs->num_syncobj_to_signal; i++) {
+ for (unsigned i = 0; i < num_syncobj_to_signal; i++) {
struct amdgpu_fence *fence =
- (struct amdgpu_fence*)cs->syncobj_to_signal[i];
+ (struct amdgpu_fence*)cs->syncobj_to_signal.list[i];
assert(amdgpu_fence_is_syncobj(fence));
sem_chunk[i].handle = fence->syncobj;
chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_SYNCOBJ_OUT;
chunks[num_chunks].length_dw = sizeof(sem_chunk[0]) / 4
- * cs->num_syncobj_to_signal;
+ * num_syncobj_to_signal;
chunks[num_chunks].chunk_data = (uintptr_t)sem_chunk;
num_chunks++;
}
+ /* Fence */
+ if (has_user_fence) {
+ chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_FENCE;
+ chunks[num_chunks].length_dw = sizeof(struct drm_amdgpu_cs_chunk_fence) / 4;
+ chunks[num_chunks].chunk_data = (uintptr_t)&acs->fence_chunk;
+ num_chunks++;
+ }
+
+ /* IB */
+ if (cs->ib[IB_PREAMBLE].ib_bytes) {
+ chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_IB;
+ chunks[num_chunks].length_dw = sizeof(struct drm_amdgpu_cs_chunk_ib) / 4;
+ chunks[num_chunks].chunk_data = (uintptr_t)&cs->ib[IB_PREAMBLE];
+ num_chunks++;
+ }
+
+ /* IB */
+ cs->ib[IB_MAIN].ib_bytes *= 4; /* Convert from dwords to bytes. */
+ chunks[num_chunks].chunk_id = AMDGPU_CHUNK_ID_IB;
+ chunks[num_chunks].length_dw = sizeof(struct drm_amdgpu_cs_chunk_ib) / 4;
+ chunks[num_chunks].chunk_data = (uintptr_t)&cs->ib[IB_MAIN];
+ num_chunks++;
+
+ if (ws->secure && cs->secure) {
+ cs->ib[IB_PREAMBLE].flags |= AMDGPU_IB_FLAGS_SECURE;
+ cs->ib[IB_MAIN].flags |= AMDGPU_IB_FLAGS_SECURE;
+ } else {
+ cs->ib[IB_PREAMBLE].flags &= ~AMDGPU_IB_FLAGS_SECURE;
+ cs->ib[IB_MAIN].flags &= ~AMDGPU_IB_FLAGS_SECURE;
+ }
+
assert(num_chunks <= ARRAY_SIZE(chunks));
- r = amdgpu_cs_submit_raw(ws->dev, acs->ctx->ctx, bo_list,
- num_chunks, chunks, &seq_no);
+ r = amdgpu_cs_submit_raw2(ws->dev, acs->ctx->ctx, bo_list,
+ num_chunks, chunks, &seq_no);
}
+finalize:
- cs->error_code = r;
if (r) {
if (r == -ENOMEM)
fprintf(stderr, "amdgpu: Not enough memory for command submission.\n");
fprintf(stderr, "amdgpu: The CS has been rejected, "
"see dmesg for more information (%i).\n", r);
- amdgpu_fence_signalled(cs->fence);
-
acs->ctx->num_rejected_cs++;
ws->num_total_rejected_cs++;
} else {
/* Cleanup. */
if (bo_list)
- amdgpu_bo_list_destroy(bo_list);
+ amdgpu_bo_list_destroy_raw(ws->dev, bo_list);
cleanup:
+ /* If there was an error, signal the fence, because it won't be signalled
+ * by the hardware. */
+ if (r)
+ amdgpu_fence_signalled(cs->fence);
+
+ cs->error_code = r;
+
for (i = 0; i < cs->num_real_buffers; i++)
p_atomic_dec(&cs->real_buffers[i].bo->num_active_ioctls);
for (i = 0; i < cs->num_slab_buffers; i++)
struct amdgpu_cs *cs = amdgpu_cs(rcs);
struct amdgpu_winsys *ws = cs->ctx->ws;
int error_code = 0;
+ uint32_t ib_pad_dw_mask = ws->info.ib_pad_dw_mask[cs->ring_type];
- rcs->current.max_dw += amdgpu_cs_epilog_dws(cs->ring_type);
+ rcs->current.max_dw += amdgpu_cs_epilog_dws(cs);
+ /* Pad the IB according to the mask. */
switch (cs->ring_type) {
case RING_DMA:
- /* pad DMA ring to 8 DWs */
- if (ws->info.chip_class <= SI) {
- while (rcs->current.cdw & 7)
+ if (ws->info.chip_class <= GFX6) {
+ while (rcs->current.cdw & ib_pad_dw_mask)
radeon_emit(rcs, 0xf0000000); /* NOP packet */
} else {
- while (rcs->current.cdw & 7)
+ while (rcs->current.cdw & ib_pad_dw_mask)
radeon_emit(rcs, 0x00000000); /* NOP packet */
}
break;
case RING_GFX:
case RING_COMPUTE:
- /* pad GFX ring to 8 DWs to meet CP fetch alignment requirements */
if (ws->info.gfx_ib_pad_with_type2) {
- while (rcs->current.cdw & 7)
- radeon_emit(rcs, 0x80000000); /* type2 nop packet */
+ while (rcs->current.cdw & ib_pad_dw_mask)
+ radeon_emit(rcs, PKT2_NOP_PAD);
} else {
- while (rcs->current.cdw & 7)
- radeon_emit(rcs, 0xffff1000); /* type3 nop packet */
+ while (rcs->current.cdw & ib_pad_dw_mask)
+ radeon_emit(rcs, PKT3_NOP_PAD);
}
if (cs->ring_type == RING_GFX)
ws->gfx_ib_size_counter += (rcs->prev_dw + rcs->current.cdw) * 4;
+
+ /* Also pad secondary IBs. */
+ if (cs->compute_ib.ib_mapped) {
+ while (cs->compute_ib.base.current.cdw & ib_pad_dw_mask)
+ radeon_emit(&cs->compute_ib.base, PKT3_NOP_PAD);
+ }
break;
case RING_UVD:
case RING_UVD_ENC:
- while (rcs->current.cdw & 15)
+ while (rcs->current.cdw & ib_pad_dw_mask)
radeon_emit(rcs, 0x80000000); /* type2 nop packet */
break;
+ case RING_VCN_JPEG:
+ if (rcs->current.cdw % 2)
+ assert(0);
+ while (rcs->current.cdw & ib_pad_dw_mask) {
+ radeon_emit(rcs, 0x60000000); /* nop packet */
+ radeon_emit(rcs, 0x00000000);
+ }
+ break;
case RING_VCN_DEC:
- while (rcs->current.cdw & 15)
+ while (rcs->current.cdw & ib_pad_dw_mask)
radeon_emit(rcs, 0x81ff); /* nop packet */
break;
default:
/* Set IB sizes. */
amdgpu_ib_finalize(ws, &cs->main);
+ if (cs->compute_ib.ib_mapped)
+ amdgpu_ib_finalize(ws, &cs->compute_ib);
+
/* Create a fence. */
amdgpu_fence_reference(&cur->fence, NULL);
if (cs->next_fence) {
/* Submit. */
util_queue_add_job(&ws->cs_queue, cs, &cs->flush_completed,
- amdgpu_cs_submit_ib, NULL);
+ amdgpu_cs_submit_ib, NULL, 0);
/* The submission has been queued, unlock the fence now. */
simple_mtx_unlock(&ws->bo_fence_lock);
amdgpu_cs_context_cleanup(cs->csc);
}
- amdgpu_get_new_ib(&ws->base, cs, IB_MAIN);
+ amdgpu_get_new_ib(ws, cs, IB_MAIN);
+ if (cs->compute_ib.ib_mapped)
+ amdgpu_get_new_ib(ws, cs, IB_PARALLEL_COMPUTE);
+
+ if (cs->preamble_ib_bo) {
+ amdgpu_cs_add_buffer(rcs, cs->preamble_ib_bo, RADEON_USAGE_READ, 0,
+ RADEON_PRIO_IB1);
+ }
cs->main.base.used_gart = 0;
cs->main.base.used_vram = 0;
amdgpu_cs_sync_flush(rcs);
util_queue_fence_destroy(&cs->flush_completed);
p_atomic_dec(&cs->ctx->ws->num_cs);
+ pb_reference(&cs->preamble_ib_bo, NULL);
pb_reference(&cs->main.big_ib_buffer, NULL);
FREE(cs->main.base.prev);
+ pb_reference(&cs->compute_ib.big_ib_buffer, NULL);
+ FREE(cs->compute_ib.base.prev);
amdgpu_destroy_cs_context(&cs->csc1);
amdgpu_destroy_cs_context(&cs->csc2);
amdgpu_fence_reference(&cs->next_fence, NULL);
return amdgpu_bo_is_referenced_by_cs_with_usage(cs, bo, usage);
}
-void amdgpu_cs_init_functions(struct amdgpu_winsys *ws)
+void amdgpu_cs_init_functions(struct amdgpu_screen_winsys *ws)
{
ws->base.ctx_create = amdgpu_ctx_create;
ws->base.ctx_destroy = amdgpu_ctx_destroy;
ws->base.ctx_query_reset_status = amdgpu_ctx_query_reset_status;
ws->base.cs_create = amdgpu_cs_create;
+ ws->base.cs_add_parallel_compute_ib = amdgpu_cs_add_parallel_compute_ib;
+ ws->base.cs_setup_preemption = amdgpu_cs_setup_preemption;
ws->base.cs_destroy = amdgpu_cs_destroy;
ws->base.cs_add_buffer = amdgpu_cs_add_buffer;
ws->base.cs_validate = amdgpu_cs_validate;