Module: Mesa Branch: main Commit: 3d3641e446794e3ebfe07de81b37c4d60eea26f0 URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=3d3641e446794e3ebfe07de81b37c4d60eea26f0
Author: Faith Ekstrand <[email protected]> Date: Tue Oct 10 10:26:19 2023 -0500 nvk: Emit MME_DMA_SYSMEMBAR before indirect draw/dispatch This fixes issues where we may read stale data from other parts of the GPU when we go to do an indirect draw fetch. Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25653> --- src/nouveau/vulkan/nvk_cmd_dispatch.c | 3 ++- src/nouveau/vulkan/nvk_cmd_draw.c | 15 ++++++++++----- 2 files changed, 12 insertions(+), 6 deletions(-) diff --git a/src/nouveau/vulkan/nvk_cmd_dispatch.c b/src/nouveau/vulkan/nvk_cmd_dispatch.c index 26cc86eb1c6..05a2c734a92 100644 --- a/src/nouveau/vulkan/nvk_cmd_dispatch.c +++ b/src/nouveau/vulkan/nvk_cmd_dispatch.c @@ -394,8 +394,9 @@ nvk_CmdDispatchIndirect(VkCommandBuffer commandBuffer, if (unlikely(qmd_addr == 0)) return; - struct nv_push *p = nvk_cmd_buffer_push(cmd, 16); + struct nv_push *p = nvk_cmd_buffer_push(cmd, 18); + P_IMMD(p, NVC597, MME_DMA_SYSMEMBAR, 0); P_IMMD(p, NVC597, SET_MME_DATA_FIFO_CONFIG, FIFO_SIZE_SIZE_4KB); P_1INC(p, NV9097, CALL_MME_MACRO(NVK_MME_DISPATCH_INDIRECT)); P_INLINE_DATA(p, nvk_compute_local_size(cmd)); diff --git a/src/nouveau/vulkan/nvk_cmd_draw.c b/src/nouveau/vulkan/nvk_cmd_draw.c index 6a9ee86233c..43e68bfb241 100644 --- a/src/nouveau/vulkan/nvk_cmd_draw.c +++ b/src/nouveau/vulkan/nvk_cmd_draw.c @@ -2101,7 +2101,8 @@ nvk_CmdDrawIndirect(VkCommandBuffer commandBuffer, }); if (nvk_cmd_buffer_3d_cls(cmd) >= TURING_A) { - struct nv_push *p = nvk_cmd_buffer_push(cmd, 8); + struct nv_push *p = nvk_cmd_buffer_push(cmd, 10); + P_IMMD(p, NVC597, MME_DMA_SYSMEMBAR, 0); P_IMMD(p, NVC597, SET_MME_DATA_FIFO_CONFIG, FIFO_SIZE_SIZE_4KB); P_1INC(p, NV9097, CALL_MME_MACRO(NVK_MME_DRAW_INDIRECT)); P_INLINE_DATA(p, begin); @@ -2218,7 +2219,8 @@ nvk_CmdDrawIndexedIndirect(VkCommandBuffer commandBuffer, }); if (nvk_cmd_buffer_3d_cls(cmd) >= TURING_A) { - struct nv_push *p = nvk_cmd_buffer_push(cmd, 8); + struct nv_push *p = nvk_cmd_buffer_push(cmd, 10); + P_IMMD(p, NVC597, MME_DMA_SYSMEMBAR, 0); P_IMMD(p, NVC597, SET_MME_DATA_FIFO_CONFIG, FIFO_SIZE_SIZE_4KB); P_1INC(p, NV9097, CALL_MME_MACRO(NVK_MME_DRAW_INDEXED_INDIRECT)); P_INLINE_DATA(p, begin); @@ -2315,7 +2317,8 @@ nvk_CmdDrawIndirectCount(VkCommandBuffer commandBuffer, .split_mode = SPLIT_MODE_NORMAL_BEGIN_NORMAL_END, }); - struct nv_push *p = nvk_cmd_buffer_push(cmd, 10); + struct nv_push *p = nvk_cmd_buffer_push(cmd, 12); + P_IMMD(p, NVC597, MME_DMA_SYSMEMBAR, 0); P_IMMD(p, NVC597, SET_MME_DATA_FIFO_CONFIG, FIFO_SIZE_SIZE_4KB); P_1INC(p, NV9097, CALL_MME_MACRO(NVK_MME_DRAW_INDIRECT_COUNT)); P_INLINE_DATA(p, begin); @@ -2392,7 +2395,8 @@ nvk_CmdDrawIndexedIndirectCount(VkCommandBuffer commandBuffer, .split_mode = SPLIT_MODE_NORMAL_BEGIN_NORMAL_END, }); - struct nv_push *p = nvk_cmd_buffer_push(cmd, 10); + struct nv_push *p = nvk_cmd_buffer_push(cmd, 12); + P_IMMD(p, NVC597, MME_DMA_SYSMEMBAR, 0); P_IMMD(p, NVC597, SET_MME_DATA_FIFO_CONFIG, FIFO_SIZE_SIZE_4KB); P_1INC(p, NV9097, CALL_MME_MACRO(NVK_MME_DRAW_INDEXED_INDIRECT_COUNT)); P_INLINE_DATA(p, begin); @@ -2508,9 +2512,10 @@ nvk_CmdDrawIndirectByteCountEXT(VkCommandBuffer commandBuffer, }); if (nvk_cmd_buffer_3d_cls(cmd) >= TURING_A) { - struct nv_push *p = nvk_cmd_buffer_push(cmd, 12); + struct nv_push *p = nvk_cmd_buffer_push(cmd, 14); P_IMMD(p, NV9097, SET_DRAW_AUTO_START, counterOffset); P_IMMD(p, NV9097, SET_DRAW_AUTO_STRIDE, vertexStride); + P_IMMD(p, NVC597, MME_DMA_SYSMEMBAR, 0); P_IMMD(p, NVC597, SET_MME_DATA_FIFO_CONFIG, FIFO_SIZE_SIZE_4KB); P_1INC(p, NV9097, CALL_MME_MACRO(NVK_MME_XFB_DRAW_INDIRECT));
