[Mesa-dev] [PATCH 01/13] anv: Use blorp for VkCmdFillBuffer
Nanley Chery
nanleychery at gmail.com
Mon Sep 26 19:26:43 UTC 2016
On Mon, Sep 26, 2016 at 12:12:32PM -0700, Jason Ekstrand wrote:
> On Sep 26, 2016 11:16 AM, "Nanley Chery" <nanleychery at gmail.com> wrote:
> >
> > On Sun, Sep 25, 2016 at 09:59:00AM -0700, Jason Ekstrand wrote:
> > > Signed-off-by: Jason Ekstrand <jason at jlekstrand.net>
> > > ---
> > > src/intel/vulkan/anv_blorp.c | 106
> +++++++++++++++++++++++++++++----
> > > src/intel/vulkan/anv_meta_clear.c | 120
> --------------------------------------
> > > 2 files changed, 96 insertions(+), 130 deletions(-)
> > >
> > > diff --git a/src/intel/vulkan/anv_blorp.c b/src/intel/vulkan/anv_blorp.c
> > > index cb61070..f5a6c40 100644
> > > --- a/src/intel/vulkan/anv_blorp.c
> > > +++ b/src/intel/vulkan/anv_blorp.c
> > > @@ -480,6 +480,20 @@ void anv_CmdBlitImage(
> > > blorp_batch_finish(&batch);
> > > }
> > >
> > > +static enum isl_format
> > > +isl_format_for_size(unsigned size_B)
> > > +{
> > > + switch (size_B) {
> > > + case 1: return ISL_FORMAT_R8_UINT;
> > > + case 2: return ISL_FORMAT_R8G8_UINT;
> > > + case 4: return ISL_FORMAT_R8G8B8A8_UINT;
> > > + case 8: return ISL_FORMAT_R16G16B16A16_UINT;
> > > + case 16: return ISL_FORMAT_R32G32B32A32_UINT;
> > > + default:
> > > + unreachable("Not a power-of-two format size");
> > > + }
> > > +}
> > > +
> > > static void
> > > do_buffer_copy(struct blorp_batch *batch,
> > > struct anv_bo *src, uint64_t src_offset,
> > > @@ -491,16 +505,7 @@ do_buffer_copy(struct blorp_batch *batch,
> > > /* The actual format we pick doesn't matter as blorp will throw it
> away.
> > > * The only thing that actually matters is the size.
> > > */
> > > - enum isl_format format;
> > > - switch (block_size) {
> > > - case 1: format = ISL_FORMAT_R8_UINT; break;
> > > - case 2: format = ISL_FORMAT_R8G8_UINT; break;
> > > - case 4: format = ISL_FORMAT_R8G8B8A8_UNORM; break;
> > > - case 8: format = ISL_FORMAT_R16G16B16A16_UNORM; break;
> > > - case 16: format = ISL_FORMAT_R32G32B32A32_UINT; break;
> > > - default:
> > > - unreachable("Not a power-of-two format size");
> > > - }
> > > + enum isl_format format = isl_format_for_size(block_size);
> > >
> > > struct isl_surf surf;
> > > isl_surf_init(&device->isl_dev, &surf,
> > > @@ -667,6 +672,87 @@ void anv_CmdUpdateBuffer(
> > > blorp_batch_finish(&batch);
> > > }
> > >
> > > +void anv_CmdFillBuffer(
> > > + VkCommandBuffer commandBuffer,
> > > + VkBuffer dstBuffer,
> > > + VkDeviceSize dstOffset,
> > > + VkDeviceSize fillSize,
> > > + uint32_t data)
> > > +{
> > > + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
> > > + ANV_FROM_HANDLE(anv_buffer, dst_buffer, dstBuffer);
> > > + struct blorp_surf surf;
> > > + struct isl_surf isl_surf;
> > > +
> > > + struct blorp_batch batch;
> > > + blorp_batch_init(&cmd_buffer->device->blorp, &batch, cmd_buffer);
> > > +
> > > + if (fillSize == VK_WHOLE_SIZE) {
> > > + fillSize = dst_buffer->size - dstOffset;
> > > + /* Make sure fillSize is a multiple of 4 */
> > > + fillSize &= ~3ull;
> > > + }
> > > +
> > > + /* First, we compute the biggest format that can be used with the
> > > + * given offsets and size.
> > > + */
> > > + int bs = 16;
> > > + bs = gcd_pow2_u64(bs, dstOffset);
> > > + bs = gcd_pow2_u64(bs, fillSize);
> > > + enum isl_format isl_format = isl_format_for_size(bs);
> > > +
> > > + union isl_color_value color = {
> > > + .u32 = { data, data, data, data },
> > > + };
> > > +
> > > + const uint64_t max_fill_size = MAX_SURFACE_DIM * MAX_SURFACE_DIM *
> bs;
> > > + while (fillSize > max_fill_size) {
> > ^
> > This should be '>='.
>
> Sure. Both work but >= is a bit clearer. Fixed locally.
>
I don't see how both could work. Wouldn't the assertion for height below fail
if fillSize == max_fill_size?
> > > + get_blorp_surf_for_anv_buffer(cmd_buffer->device,
> > > + dst_buffer, dstOffset,
> > > + MAX_SURFACE_DIM, MAX_SURFACE_DIM,
> > > + MAX_SURFACE_DIM * bs, isl_format,
> > > + &surf, &isl_surf);
> > > +
> > > + blorp_clear(&batch, &surf, isl_format, ISL_SWIZZLE_IDENTITY,
> > > + 0, 0, 1, 0, 0, MAX_SURFACE_DIM, MAX_SURFACE_DIM,
> > > + color, NULL);
> > > + fillSize -= max_fill_size;
> > > + dstOffset += max_fill_size;
> > > + }
> > > +
> > > + uint64_t height = fillSize / (MAX_SURFACE_DIM * bs);
> > > + assert(height < MAX_SURFACE_DIM);
> > > + if (height != 0) {
> > > + const uint64_t rect_fill_size = height * MAX_SURFACE_DIM * bs;
> > > + get_blorp_surf_for_anv_buffer(cmd_buffer->device,
> > > + dst_buffer, dstOffset,
> > > + MAX_SURFACE_DIM, height,
> > > + MAX_SURFACE_DIM * bs, isl_format,
> > > + &surf, &isl_surf);
> > > +
> > > + blorp_clear(&batch, &surf, isl_format, ISL_SWIZZLE_IDENTITY,
> > > + 0, 0, 1, 0, 0, MAX_SURFACE_DIM, height,
> > > + color, NULL);
> > > + fillSize -= rect_fill_size;
> > > + dstOffset += rect_fill_size;
> > > + }
> > > +
> > > + if (fillSize != 0) {
> > > + const uint32_t width = fillSize / bs;
> > > + get_blorp_surf_for_anv_buffer(cmd_buffer->device,
> > > + dst_buffer, dstOffset,
> > > + width, 1,
> > > + width * bs, isl_format,
> > ^
> > fillSize could be used instead.
>
> Sure. I mostly did it that way to match the other two callers of
> get_blorp_surf_for_anv_buffer.
>
Gotcha.
-Nanley
> > -Nanley
> >
> > > + &surf, &isl_surf);
> > > +
> > > + blorp_clear(&batch, &surf, isl_format, ISL_SWIZZLE_IDENTITY,
> > > + 0, 0, 1, 0, 0, width, 1,
> > > + color, NULL);
> > > + }
> > > +
> > > + blorp_batch_finish(&batch);
> > > +}
> > > +
> > > void anv_CmdClearColorImage(
> > > VkCommandBuffer commandBuffer,
> > > VkImage _image,
> > > diff --git a/src/intel/vulkan/anv_meta_clear.c
> b/src/intel/vulkan/anv_meta_clear.c
> > > index fd0797f..5579454 100644
> > > --- a/src/intel/vulkan/anv_meta_clear.c
> > > +++ b/src/intel/vulkan/anv_meta_clear.c
> > > @@ -944,123 +944,3 @@ void anv_CmdClearAttachments(
> > >
> > > meta_clear_end(&saved_state, cmd_buffer);
> > > }
> > > -
> > > -static void
> > > -do_buffer_fill(struct anv_cmd_buffer *cmd_buffer,
> > > - struct anv_bo *dest, uint64_t dest_offset,
> > > - int width, int height, VkFormat fill_format, uint32_t
> data)
> > > -{
> > > - VkDevice vk_device = anv_device_to_handle(cmd_buffer->device);
> > > -
> > > - VkImageCreateInfo image_info = {
> > > - .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
> > > - .imageType = VK_IMAGE_TYPE_2D,
> > > - .format = fill_format,
> > > - .extent = {
> > > - .width = width,
> > > - .height = height,
> > > - .depth = 1,
> > > - },
> > > - .mipLevels = 1,
> > > - .arrayLayers = 1,
> > > - .samples = 1,
> > > - .tiling = VK_IMAGE_TILING_LINEAR,
> > > - .usage = VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT,
> > > - .flags = 0,
> > > - };
> > > -
> > > - VkImage dest_image;
> > > - image_info.usage = VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT;
> > > - anv_CreateImage(vk_device, &image_info,
> > > - &cmd_buffer->pool->alloc, &dest_image);
> > > -
> > > - /* We could use a vk call to bind memory, but that would require
> > > - * creating a dummy memory object etc. so there's really no point.
> > > - */
> > > - anv_image_from_handle(dest_image)->bo = dest;
> > > - anv_image_from_handle(dest_image)->offset = dest_offset;
> > > -
> > > - const VkClearValue clear_value = {
> > > - .color = {
> > > - .uint32 = { data, data, data, data }
> > > - }
> > > - };
> > > -
> > > - const VkImageSubresourceRange range = {
> > > - .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
> > > - .baseMipLevel = 0,
> > > - .levelCount = 1,
> > > - .baseArrayLayer = 0,
> > > - .layerCount = 1,
> > > - };
> > > -
> > > - anv_cmd_clear_image(cmd_buffer, anv_image_from_handle(dest_image),
> > > - VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
> > > - clear_value, 1, &range);
> > > -}
> > > -
> > > -void anv_CmdFillBuffer(
> > > - VkCommandBuffer commandBuffer,
> > > - VkBuffer dstBuffer,
> > > - VkDeviceSize dstOffset,
> > > - VkDeviceSize fillSize,
> > > - uint32_t data)
> > > -{
> > > - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
> > > - ANV_FROM_HANDLE(anv_buffer, dst_buffer, dstBuffer);
> > > - struct anv_meta_saved_state saved_state;
> > > -
> > > - meta_clear_begin(&saved_state, cmd_buffer);
> > > -
> > > - if (fillSize == VK_WHOLE_SIZE) {
> > > - fillSize = dst_buffer->size - dstOffset;
> > > - /* Make sure fillSize is a multiple of 4 */
> > > - fillSize -= fillSize & 3;
> > > - }
> > > -
> > > - VkFormat format;
> > > - int bs;
> > > - if ((fillSize & 15) == 0 && (dstOffset & 15) == 0) {
> > > - format = VK_FORMAT_R32G32B32A32_UINT;
> > > - bs = 16;
> > > - } else if ((fillSize & 7) == 0 && (dstOffset & 15) == 0) {
> > > - format = VK_FORMAT_R32G32_UINT;
> > > - bs = 8;
> > > - } else {
> > > - assert((fillSize & 3) == 0 && (dstOffset & 3) == 0);
> > > - format = VK_FORMAT_R32_UINT;
> > > - bs = 4;
> > > - }
> > > -
> > > - /* This is maximum possible width/height our HW can handle */
> > > - const uint64_t max_surface_dim = 1 << 14;
> > > -
> > > - /* First, we make a bunch of max-sized copies */
> > > - const uint64_t max_fill_size = max_surface_dim * max_surface_dim *
> bs;
> > > - while (fillSize > max_fill_size) {
> > > - do_buffer_fill(cmd_buffer, dst_buffer->bo,
> > > - dst_buffer->offset + dstOffset,
> > > - max_surface_dim, max_surface_dim, format, data);
> > > - fillSize -= max_fill_size;
> > > - dstOffset += max_fill_size;
> > > - }
> > > -
> > > - uint64_t height = fillSize / (max_surface_dim * bs);
> > > - assert(height < max_surface_dim);
> > > - if (height != 0) {
> > > - const uint64_t rect_fill_size = height * max_surface_dim * bs;
> > > - do_buffer_fill(cmd_buffer, dst_buffer->bo,
> > > - dst_buffer->offset + dstOffset,
> > > - max_surface_dim, height, format, data);
> > > - fillSize -= rect_fill_size;
> > > - dstOffset += rect_fill_size;
> > > - }
> > > -
> > > - if (fillSize != 0) {
> > > - do_buffer_fill(cmd_buffer, dst_buffer->bo,
> > > - dst_buffer->offset + dstOffset,
> > > - fillSize / bs, 1, format, data);
> > > - }
> > > -
> > > - meta_clear_end(&saved_state, cmd_buffer);
> > > -}
> > > --
> > > 2.5.0.400.gff86faf
> > >
> > > _______________________________________________
> > > mesa-dev mailing list
> > > mesa-dev at lists.freedesktop.org
> > > https://lists.freedesktop.org/mailman/listinfo/mesa-dev
More information about the mesa-dev
mailing list