1
0
mirror of https://github.com/Yours3lf/rpi-vk-driver.git synced 2024-12-01 13:24:20 +01:00

cmdclear now only flags image for clearing

actual clear happens in pipelinebarrier
This commit is contained in:
Unknown 2018-08-25 18:10:07 +01:00
parent 356c7d3277
commit 0c3e0f798b
2 changed files with 160 additions and 120 deletions

View File

@ -1376,7 +1376,159 @@ VKAPI_ATTR void VKAPI_CALL vkCmdPipelineBarrier(
{
assert(commandBuffer);
//TODO
//TODO pipeline stage flags
//VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT
//VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT
//VK_PIPELINE_STAGE_VERTEX_INPUT_BIT
//VK_PIPELINE_STAGE_VERTEX_SHADER_BIT
//VK_PIPELINE_STAGE_TESSELLATION_CONTROL_SHADER_BIT
//VK_PIPELINE_STAGE_TESSELLATION_EVALUATION_SHADER_BIT
//VK_PIPELINE_STAGE_GEOMETRY_SHADER_BIT
//VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT
//VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT
//VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT
//VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT
//VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT
//VK_PIPELINE_STAGE_TRANSFER_BIT
//VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT
//VK_PIPELINE_STAGE_HOST_BIT
//VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT
//VK_PIPELINE_STAGE_ALL_COMMANDS_BIT
//TODO dependency flags
//VK_DEPENDENCY_BY_REGION_BIT,
//VK_DEPENDENCY_DEVICE_GROUP_BIT,
//VK_DEPENDENCY_VIEW_LOCAL_BIT
//TODO access flags
//VK_ACCESS_INDIRECT_COMMAND_READ_BIT
//VK_ACCESS_INDEX_READ_BIT
//VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT
//VK_ACCESS_UNIFORM_READ_BIT
//VK_ACCESS_INPUT_ATTACHMENT_READ_BIT
//VK_ACCESS_SHADER_READ_BIT
//VK_ACCESS_SHADER_WRITE_BIT
//VK_ACCESS_COLOR_ATTACHMENT_READ_BIT
//VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT
//VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT
//VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT
//VK_ACCESS_TRANSFER_READ_BIT
//VK_ACCESS_TRANSFER_WRITE_BIT
//VK_ACCESS_HOST_READ_BIT
//VK_ACCESS_HOST_WRITE_BIT
//VK_ACCESS_MEMORY_READ_BIT
//VK_ACCESS_MEMORY_WRITE_BIT
//VK_ACCESS_COMMAND_PROCESS_READ_BIT_NVX
//VK_ACCESS_COMMAND_PROCESS_WRITE_BIT_NVX
//TODO Layout transition flags
//VK_IMAGE_LAYOUT_UNDEFINED
//VK_IMAGE_LAYOUT_GENERAL
//VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL
//VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL
//VK_IMAGE_LAYOUT_DEPTH_STENCIL_READ_ONLY_OPTIMAL
//VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL
//VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL
//VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
//VK_IMAGE_LAYOUT_PREINITIALIZED
//VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_STENCIL_ATTACHMENT_OPTIMAL
//VK_IMAGE_LAYOUT_DEPTH_ATTACHMENT_STENCIL_READ_ONLY_OPTIMAL
//VK_IMAGE_LAYOUT_PRESENT_SRC_KHR
//VK_IMAGE_LAYOUT_SHARED_PRESENT_KHR
for(int c = 0; c < memoryBarrierCount; ++c)
{
//TODO
}
for(int c = 0; c < bufferMemoryBarrierCount; ++c)
{
//TODO
}
for(int c = 0; c < imageMemoryBarrierCount; ++c)
{
_image* i = pImageMemoryBarriers[c].image;
assert(i->layout == pImageMemoryBarriers[c].oldLayout || i->layout == VK_IMAGE_LAYOUT_UNDEFINED);
if(srcStageMask & VK_PIPELINE_STAGE_TRANSFER_BIT &&
pImageMemoryBarriers[c].srcAccessMask & VK_ACCESS_TRANSFER_WRITE_BIT &&
i->needToClear)
{
//insert CRs to clear the image
assert(i->layout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
clFit(commandBuffer, &commandBuffer->binCl, V3D21_TILE_BINNING_MODE_CONFIGURATION_length);
clInsertTileBinningModeConfiguration(&commandBuffer->binCl,
0, 0, 0, 0,
getFormatBpp(i->format) == 64, //64 bit color mode
i->samples > 1, //msaa
i->width, i->height, 0, 0, 0);
//START_TILE_BINNING resets the statechange counters in the hardware,
//which are what is used when a primitive is binned to a tile to
//figure out what new state packets need to be written to that tile's
//command list.
clFit(commandBuffer, &commandBuffer->binCl, V3D21_START_TILE_BINNING_length);
clInsertStartTileBinning(&commandBuffer->binCl);
//Reset the current compressed primitives format. This gets modified
//by VC4_PACKET_GL_INDEXED_PRIMITIVE and
//VC4_PACKET_GL_ARRAY_PRIMITIVE, so it needs to be reset at the start
//of every tile.
clFit(commandBuffer, &commandBuffer->binCl, V3D21_PRIMITIVE_LIST_FORMAT_length);
clInsertPrimitiveListFormat(&commandBuffer->binCl,
1, //16 bit
2); //tris
clFit(commandBuffer, &commandBuffer->handlesCl, 4);
uint32_t idx = clGetHandleIndex(&commandBuffer->handlesCl, i->handle);
commandBuffer->submitCl.color_write.hindex = idx;
commandBuffer->submitCl.color_write.offset = 0;
commandBuffer->submitCl.color_write.flags = 0;
//TODO format
commandBuffer->submitCl.color_write.bits =
VC4_SET_FIELD(VC4_RENDER_CONFIG_FORMAT_RGBA8888, VC4_RENDER_CONFIG_FORMAT) |
VC4_SET_FIELD(i->tiling, VC4_RENDER_CONFIG_MEMORY_FORMAT);
commandBuffer->submitCl.clear_color[0] = i->clearColor[0];
commandBuffer->submitCl.clear_color[1] = i->clearColor[1];
//TODO ranges
commandBuffer->submitCl.min_x_tile = 0;
commandBuffer->submitCl.min_y_tile = 0;
uint32_t tileSizeW = 64;
uint32_t tileSizeH = 64;
if(i->samples > 1)
{
tileSizeW >>= 1;
tileSizeH >>= 1;
}
if(getFormatBpp(i->format) == 64)
{
tileSizeH >>= 1;
}
uint32_t widthInTiles = divRoundUp(i->width, tileSizeW);
uint32_t heightInTiles = divRoundUp(i->height, tileSizeH);
commandBuffer->submitCl.max_x_tile = widthInTiles - 1;
commandBuffer->submitCl.max_y_tile = heightInTiles - 1;
commandBuffer->submitCl.width = i->width;
commandBuffer->submitCl.height = i->height;
commandBuffer->submitCl.flags |= VC4_SUBMIT_CL_USE_CLEAR_COLOR;
commandBuffer->submitCl.clear_z = 0; //TODO
commandBuffer->submitCl.clear_s = 0;
}
//transition to new layout
i->layout = pImageMemoryBarriers[c].newLayout;
}
}
/*
@ -1400,7 +1552,7 @@ VKAPI_ATTR void VKAPI_CALL vkCmdClearColorImage(
//actual clearing would only happen:
// -if image is rendered to (insert clear before first draw call)
// -if the image is bound for sampling (submit a CL with a clear)
// -if the command buffer is closed without any rendering (insert clear)
// -if a command buffer is submitted without any rendering (insert clear)
// -etc.
//we shouldn't clear an image if noone uses it
@ -1413,129 +1565,14 @@ VKAPI_ATTR void VKAPI_CALL vkCmdClearColorImage(
assert(commandBuffer->state == CMDBUF_STATE_RECORDING);
assert(_queueFamilyProperties[commandBuffer->cp->queueFamilyIndex].queueFlags & VK_QUEUE_GRAPHICS_BIT || _queueFamilyProperties[commandBuffer->cp->queueFamilyIndex].queueFlags & VK_QUEUE_COMPUTE_BIT);
//TODO externally sync cmdbuf, cmdpool
_image* i = image;
assert(i->usageBits & VK_IMAGE_USAGE_TRANSFER_DST_BIT);
clFit(commandBuffer, &commandBuffer->binCl, V3D21_TILE_BINNING_MODE_CONFIGURATION_length);
clInsertTileBinningModeConfiguration(&commandBuffer->binCl,
0, 0, 0, 0,
getFormatBpp(i->format) == 64, //64 bit color mode
i->samples > 1, //msaa
i->width, i->height, 0, 0, 0);
//TODO externally sync cmdbuf, cmdpool
//START_TILE_BINNING resets the statechange counters in the hardware,
//which are what is used when a primitive is binned to a tile to
//figure out what new state packets need to be written to that tile's
//command list.
clFit(commandBuffer, &commandBuffer->binCl, V3D21_START_TILE_BINNING_length);
clInsertStartTileBinning(&commandBuffer->binCl);
//Reset the current compressed primitives format. This gets modified
//by VC4_PACKET_GL_INDEXED_PRIMITIVE and
//VC4_PACKET_GL_ARRAY_PRIMITIVE, so it needs to be reset at the start
//of every tile.
clFit(commandBuffer, &commandBuffer->binCl, V3D21_PRIMITIVE_LIST_FORMAT_length);
clInsertPrimitiveListFormat(&commandBuffer->binCl,
1, //16 bit
2); //tris
clFit(commandBuffer, &commandBuffer->handlesCl, 4);
uint32_t idx = clGetHandleIndex(&commandBuffer->handlesCl, i->handle);
commandBuffer->submitCl.color_write.hindex = idx;
commandBuffer->submitCl.color_write.offset = 0;
commandBuffer->submitCl.color_write.flags = 0;
//TODO format
commandBuffer->submitCl.color_write.bits =
VC4_SET_FIELD(VC4_RENDER_CONFIG_FORMAT_RGBA8888, VC4_RENDER_CONFIG_FORMAT) |
VC4_SET_FIELD(i->tiling, VC4_RENDER_CONFIG_MEMORY_FORMAT);
//TODO msaa?
commandBuffer->submitCl.clear_color[0] =
commandBuffer->submitCl.clear_color[1] = packVec4IntoABGR8(pColor->float32);
//TODO ranges
commandBuffer->submitCl.min_x_tile = 0;
commandBuffer->submitCl.min_y_tile = 0;
uint32_t tileSizeW = 64;
uint32_t tileSizeH = 64;
if(i->samples > 1)
{
tileSizeW >>= 1;
tileSizeH >>= 1;
}
if(getFormatBpp(i->format) == 64)
{
tileSizeH >>= 1;
}
uint32_t widthInTiles = divRoundUp(i->width, tileSizeW);
uint32_t heightInTiles = divRoundUp(i->height, tileSizeH);
commandBuffer->submitCl.max_x_tile = widthInTiles - 1;
commandBuffer->submitCl.max_y_tile = heightInTiles - 1;
commandBuffer->submitCl.width = i->width;
commandBuffer->submitCl.height = i->height;
commandBuffer->submitCl.flags |= VC4_SUBMIT_CL_USE_CLEAR_COLOR;
commandBuffer->submitCl.clear_z = 0; //TODO
commandBuffer->submitCl.clear_s = 0;
//TODO primitive list format must be followed by shader state
//clFit(commandBuffer, &commandBuffer->binCl, V3D21_GL_SHADER_STATE_length);
//clInsertShaderState(&commandBuffer->binCl, 0, 0, 0);
//clFit(commandBuffer, &commandBuffer->binCl, V3D21_CLIP_WINDOW_length);
//clInsertClipWindow(&commandBuffer->binCl, i->width, i->height, 0, 0); //TODO should this be configurable?
/*
//TODO
clFit(commandBuffer, &commandBuffer->binCl, V3D21_CONFIGURATION_BITS_length);
clInsertConfigurationBits(&commandBuffer->binCl,
0,
0,
0,
V3D_COMPARE_FUNC_ALWAYS,
0,
0,
0,
0,
0,
0,
0,
1,
1);
//TODO
clFit(commandBuffer, &commandBuffer->binCl, V3D21_DEPTH_OFFSET_length);
clInsertDepthOffset(&commandBuffer->binCl, 0, 0);
clFit(commandBuffer, &commandBuffer->binCl, V3D21_POINT_SIZE_length);
clInsertPointSize(&commandBuffer->binCl, 1.0f);
clFit(commandBuffer, &commandBuffer->binCl, V3D21_LINE_WIDTH_length);
clInsertLineWidth(&commandBuffer->binCl, 1.0f);
//TODO
clFit(commandBuffer, &commandBuffer->binCl, V3D21_CLIPPER_XY_SCALING_length);
clInsertClipperXYScaling(&commandBuffer->binCl, 1.0f, 1.0f);
clFit(commandBuffer, &commandBuffer->binCl, V3D21_CLIPPER_Z_SCALE_AND_OFFSET_length);
clInsertClipperZScaleOffset(&commandBuffer->binCl, 0.0f, 1.0f);
clFit(commandBuffer, &commandBuffer->binCl, V3D21_VIEWPORT_OFFSET_length);
clInsertViewPortOffset(&commandBuffer->binCl, 0, 0);
//TODO
clFit(commandBuffer, &commandBuffer->binCl, V3D21_FLAT_SHADE_FLAGS_length);
clInsertFlatShadeFlags(&commandBuffer->binCl, 0);
//TODO I suppose this should be a submit itself?
*/
i->needToClear = 1;
i->clearColor[0] = i->clearColor[1] = packVec4IntoABGR8(pColor->float32);
}
/*

View File

@ -32,6 +32,9 @@ typedef struct VkImage_T
uint32_t format;
uint32_t imageSpace;
uint32_t tiling;
uint32_t needToClear;
uint32_t clearColor[2];
uint32_t layout;
uint32_t concurrentAccess; //TODO
uint32_t numQueueFamiliesWithAccess;
uint32_t* queueFamiliesWithAccess;