X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Ffreedreno%2Ffreedreno_context.c;h=ca3b793000d2eb159741fb8a48e48077381714cd;hb=b8fbb39ab2c962e38f6c9d668de57582faf39b70;hp=b9a1fe977457370edcc117f113b2d6e2fd047a19;hpb=f02a64dbdd2ec147167ad60357bd46d8d964290a;p=mesa.git diff --git a/src/gallium/drivers/freedreno/freedreno_context.c b/src/gallium/drivers/freedreno/freedreno_context.c index b9a1fe97745..ca3b793000d 100644 --- a/src/gallium/drivers/freedreno/freedreno_context.c +++ b/src/gallium/drivers/freedreno/freedreno_context.c @@ -1,5 +1,3 @@ -/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */ - /* * Copyright (C) 2012 Rob Clark * @@ -27,6 +25,7 @@ */ #include "freedreno_context.h" +#include "freedreno_blitter.h" #include "freedreno_draw.h" #include "freedreno_fence.h" #include "freedreno_program.h" @@ -37,40 +36,89 @@ #include "freedreno_query.h" #include "freedreno_query_hw.h" #include "freedreno_util.h" +#include "util/u_upload_mgr.h" -/* emit accumulated render cmds, needed for example if render target has - * changed, or for flush() - */ -void -fd_context_render(struct pipe_context *pctx) +static void +fd_context_flush(struct pipe_context *pctx, struct pipe_fence_handle **fencep, + unsigned flags) { struct fd_context *ctx = fd_context(pctx); - struct fd_batch *new_batch; + struct pipe_fence_handle *fence = NULL; + // TODO we want to lookup batch if it exists, but not create one if not. + struct fd_batch *batch = fd_context_batch(ctx); + + DBG("%p: flush: flags=%x\n", ctx->batch, flags); + + /* In some sequence of events, we can end up with a last_fence that is + * not an "fd" fence, which results in eglDupNativeFenceFDANDROID() + * errors. + * + */ + if (flags & PIPE_FLUSH_FENCE_FD) + fd_fence_ref(&ctx->last_fence, NULL); + + /* if no rendering since last flush, ie. app just decided it needed + * a fence, re-use the last one: + */ + if (ctx->last_fence) { + fd_fence_ref(&fence, ctx->last_fence); + goto out; + } + + if (!batch) + return; - fd_batch_flush(ctx->batch); + /* Take a ref to the batch's fence (batch can be unref'd when flushed: */ + fd_fence_ref(&fence, batch->fence); - new_batch = fd_batch_create(ctx); - util_copy_framebuffer_state(&new_batch->framebuffer, &ctx->batch->framebuffer); - fd_batch_reference(&ctx->batch, NULL); - ctx->batch = new_batch; + if (flags & PIPE_FLUSH_FENCE_FD) + batch->needs_out_fence_fd = true; + + if (!ctx->screen->reorder) { + fd_batch_flush(batch, true); + } else if (flags & PIPE_FLUSH_DEFERRED) { + fd_bc_flush_deferred(&ctx->screen->batch_cache, ctx); + } else { + fd_bc_flush(&ctx->screen->batch_cache, ctx); + } + +out: + if (fencep) + fd_fence_ref(fencep, fence); + + fd_fence_ref(&ctx->last_fence, fence); + + fd_fence_ref(&fence, NULL); } static void -fd_context_flush(struct pipe_context *pctx, struct pipe_fence_handle **fence, - unsigned flags) +fd_texture_barrier(struct pipe_context *pctx, unsigned flags) { - struct fd_batch *batch = NULL; + if (flags == PIPE_TEXTURE_BARRIER_FRAMEBUFFER) { + struct fd_context *ctx = fd_context(pctx); - fd_batch_reference(&batch, fd_context(pctx)->batch); + if (ctx->framebuffer_barrier) { + ctx->framebuffer_barrier(ctx); + return; + } + } - fd_context_render(pctx); + /* On devices that could sample from GMEM we could possibly do better. + * Or if we knew that we were doing GMEM bypass we could just emit a + * cache flush, perhaps? But we don't know if future draws would cause + * us to use GMEM, and a flush in bypass isn't the end of the world. + */ + fd_context_flush(pctx, NULL, 0); +} - if (fence) { - fd_screen_fence_ref(pctx->screen, fence, NULL); - *fence = fd_fence_create(pctx, fd_ringbuffer_timestamp(batch->gmem)); - } +static void +fd_memory_barrier(struct pipe_context *pctx, unsigned flags) +{ + if (!(flags & ~PIPE_BARRIER_UPDATE)) + return; - fd_batch_reference(&batch, NULL); + fd_context_flush(pctx, NULL, 0); + /* TODO do we need to check for persistently mapped buffers and fd_bo_cpu_prep()?? */ } /** @@ -81,13 +129,23 @@ static void fd_emit_string_marker(struct pipe_context *pctx, const char *string, int len) { struct fd_context *ctx = fd_context(pctx); - struct fd_ringbuffer *ring = ctx->batch->draw; + struct fd_ringbuffer *ring; const uint32_t *buf = (const void *)string; + if (!ctx->batch) + return; + + ctx->batch->needs_flush = true; + + ring = ctx->batch->draw; + /* max packet size is 0x3fff dwords: */ len = MIN2(len, 0x3fff * 4); - OUT_PKT3(ring, CP_NOP, align(len, 4) / 4); + if (ctx->screen->gpu_id >= 500) + OUT_PKT7(ring, CP_NOP, align(len, 4) / 4); + else + OUT_PKT3(ring, CP_NOP, align(len, 4) / 4); while (len >= 4) { OUT_RING(ring, *buf); buf++; @@ -110,29 +168,47 @@ fd_context_destroy(struct pipe_context *pctx) DBG(""); + fd_fence_ref(&ctx->last_fence, NULL); + + if (ctx->screen->reorder && util_queue_is_initialized(&ctx->flush_queue)) + util_queue_destroy(&ctx->flush_queue); + + util_copy_framebuffer_state(&ctx->framebuffer, NULL); + fd_batch_reference(&ctx->batch, NULL); /* unref current batch */ + fd_bc_invalidate_context(ctx); + fd_prog_fini(pctx); - fd_hw_query_fini(pctx); if (ctx->blitter) util_blitter_destroy(ctx->blitter); + if (pctx->stream_uploader) + u_upload_destroy(pctx->stream_uploader); + + if (ctx->clear_rs_state) + pctx->delete_rasterizer_state(pctx, ctx->clear_rs_state); + if (ctx->primconvert) util_primconvert_destroy(ctx->primconvert); - util_slab_destroy(&ctx->transfer_pool); - - fd_batch_reference(&ctx->batch, NULL); /* unref current batch */ + slab_destroy_child(&ctx->transfer_pool); - for (i = 0; i < ARRAY_SIZE(ctx->pipe); i++) { - struct fd_vsc_pipe *pipe = &ctx->pipe[i]; + for (i = 0; i < ARRAY_SIZE(ctx->vsc_pipe); i++) { + struct fd_vsc_pipe *pipe = &ctx->vsc_pipe[i]; if (!pipe->bo) break; fd_bo_del(pipe->bo); } fd_device_del(ctx->dev); + fd_pipe_del(ctx->pipe); - FREE(ctx); + if (fd_mesa_debug & (FD_DBG_BSTAT | FD_DBG_MSGS)) { + printf("batch_total=%u, batch_sysmem=%u, batch_gmem=%u, batch_nondraw=%u, batch_restore=%u\n", + (uint32_t)ctx->stats.batch_total, (uint32_t)ctx->stats.batch_sysmem, + (uint32_t)ctx->stats.batch_gmem, (uint32_t)ctx->stats.batch_nondraw, + (uint32_t)ctx->stats.batch_restore); + } } static void @@ -147,15 +223,139 @@ fd_set_debug_callback(struct pipe_context *pctx, memset(&ctx->debug, 0, sizeof(ctx->debug)); } +static uint32_t +fd_get_reset_count(struct fd_context *ctx, bool per_context) +{ + uint64_t val; + enum fd_param_id param = + per_context ? FD_CTX_FAULTS : FD_GLOBAL_FAULTS; + int ret = fd_pipe_get_param(ctx->pipe, param, &val); + debug_assert(!ret); + return val; +} + +static enum pipe_reset_status +fd_get_device_reset_status(struct pipe_context *pctx) +{ + struct fd_context *ctx = fd_context(pctx); + int context_faults = fd_get_reset_count(ctx, true); + int global_faults = fd_get_reset_count(ctx, false); + enum pipe_reset_status status; + + if (context_faults != ctx->context_reset_count) { + status = PIPE_GUILTY_CONTEXT_RESET; + } else if (global_faults != ctx->global_reset_count) { + status = PIPE_INNOCENT_CONTEXT_RESET; + } else { + status = PIPE_NO_RESET; + } + + ctx->context_reset_count = context_faults; + ctx->global_reset_count = global_faults; + + return status; +} + +/* TODO we could combine a few of these small buffers (solid_vbuf, + * blit_texcoord_vbuf, and vsc_size_mem, into a single buffer and + * save a tiny bit of memory + */ + +static struct pipe_resource * +create_solid_vertexbuf(struct pipe_context *pctx) +{ + static const float init_shader_const[] = { + -1.000000, +1.000000, +1.000000, + +1.000000, -1.000000, +1.000000, + }; + struct pipe_resource *prsc = pipe_buffer_create(pctx->screen, + PIPE_BIND_CUSTOM, PIPE_USAGE_IMMUTABLE, sizeof(init_shader_const)); + pipe_buffer_write(pctx, prsc, 0, + sizeof(init_shader_const), init_shader_const); + return prsc; +} + +static struct pipe_resource * +create_blit_texcoord_vertexbuf(struct pipe_context *pctx) +{ + struct pipe_resource *prsc = pipe_buffer_create(pctx->screen, + PIPE_BIND_CUSTOM, PIPE_USAGE_DYNAMIC, 16); + return prsc; +} + +void +fd_context_setup_common_vbos(struct fd_context *ctx) +{ + struct pipe_context *pctx = &ctx->base; + + ctx->solid_vbuf = create_solid_vertexbuf(pctx); + ctx->blit_texcoord_vbuf = create_blit_texcoord_vertexbuf(pctx); + + /* setup solid_vbuf_state: */ + ctx->solid_vbuf_state.vtx = pctx->create_vertex_elements_state( + pctx, 1, (struct pipe_vertex_element[]){{ + .vertex_buffer_index = 0, + .src_offset = 0, + .src_format = PIPE_FORMAT_R32G32B32_FLOAT, + }}); + ctx->solid_vbuf_state.vertexbuf.count = 1; + ctx->solid_vbuf_state.vertexbuf.vb[0].stride = 12; + ctx->solid_vbuf_state.vertexbuf.vb[0].buffer.resource = ctx->solid_vbuf; + + /* setup blit_vbuf_state: */ + ctx->blit_vbuf_state.vtx = pctx->create_vertex_elements_state( + pctx, 2, (struct pipe_vertex_element[]){{ + .vertex_buffer_index = 0, + .src_offset = 0, + .src_format = PIPE_FORMAT_R32G32_FLOAT, + }, { + .vertex_buffer_index = 1, + .src_offset = 0, + .src_format = PIPE_FORMAT_R32G32B32_FLOAT, + }}); + ctx->blit_vbuf_state.vertexbuf.count = 2; + ctx->blit_vbuf_state.vertexbuf.vb[0].stride = 8; + ctx->blit_vbuf_state.vertexbuf.vb[0].buffer.resource = ctx->blit_texcoord_vbuf; + ctx->blit_vbuf_state.vertexbuf.vb[1].stride = 12; + ctx->blit_vbuf_state.vertexbuf.vb[1].buffer.resource = ctx->solid_vbuf; +} + +void +fd_context_cleanup_common_vbos(struct fd_context *ctx) +{ + struct pipe_context *pctx = &ctx->base; + + pctx->delete_vertex_elements_state(pctx, ctx->solid_vbuf_state.vtx); + pctx->delete_vertex_elements_state(pctx, ctx->blit_vbuf_state.vtx); + + pipe_resource_reference(&ctx->solid_vbuf, NULL); + pipe_resource_reference(&ctx->blit_texcoord_vbuf, NULL); +} + struct pipe_context * fd_context_init(struct fd_context *ctx, struct pipe_screen *pscreen, - const uint8_t *primtypes, void *priv) + const uint8_t *primtypes, void *priv, unsigned flags) { struct fd_screen *screen = fd_screen(pscreen); struct pipe_context *pctx; + unsigned prio = 1; int i; + /* lower numerical value == higher priority: */ + if (fd_mesa_debug & FD_DBG_HIPRIO) + prio = 0; + else if (flags & PIPE_CONTEXT_HIGH_PRIORITY) + prio = 0; + else if (flags & PIPE_CONTEXT_LOW_PRIORITY) + prio = 2; + ctx->screen = screen; + ctx->pipe = fd_pipe_new2(screen->dev, FD_PIPE_3D, prio); + + if (fd_device_version(screen->dev) >= FD_VERSION_ROBUSTNESS) { + ctx->context_reset_count = fd_get_reset_count(ctx, true); + ctx->global_reset_count = fd_get_reset_count(ctx, false); + } ctx->primtypes = primtypes; ctx->primtype_mask = 0; @@ -168,28 +368,33 @@ fd_context_init(struct fd_context *ctx, struct pipe_screen *pscreen, */ ctx->sample_mask = 0xffff; - ctx->stage = FD_STAGE_NULL; - pctx = &ctx->base; pctx->screen = pscreen; pctx->priv = priv; pctx->flush = fd_context_flush; pctx->emit_string_marker = fd_emit_string_marker; pctx->set_debug_callback = fd_set_debug_callback; + pctx->get_device_reset_status = fd_get_device_reset_status; + pctx->create_fence_fd = fd_create_fence_fd; + pctx->fence_server_sync = fd_fence_server_sync; + pctx->texture_barrier = fd_texture_barrier; + pctx->memory_barrier = fd_memory_barrier; + + pctx->stream_uploader = u_upload_create_default(pctx); + if (!pctx->stream_uploader) + goto fail; + pctx->const_uploader = pctx->stream_uploader; - ctx->batch = fd_batch_create(ctx); - - fd_reset_wfi(ctx); + if (!ctx->screen->reorder) + ctx->batch = fd_bc_alloc_batch(&screen->batch_cache, ctx, false); - util_slab_create(&ctx->transfer_pool, sizeof(struct fd_transfer), - 16, UTIL_SLAB_SINGLETHREADED); + slab_create_child(&ctx->transfer_pool, &screen->transfer_pool); fd_draw_init(pctx); fd_resource_context_init(pctx); fd_query_context_init(pctx); fd_texture_init(pctx); fd_state_init(pctx); - fd_hw_query_init(pctx); ctx->blitter = util_blitter_create(pctx); if (!ctx->blitter) @@ -199,6 +404,9 @@ fd_context_init(struct fd_context *ctx, struct pipe_screen *pscreen, if (!ctx->primconvert) goto fail; + list_inithead(&ctx->hw_active_queries); + list_inithead(&ctx->acc_active_queries); + return pctx; fail: