[Mesa-dev] [PATCH v2] radv: Handle command buffers that need scratch memory.

Dave Airlie airlied at gmail.com
Mon Jan 30 00:11:53 UTC 2017


On 30 January 2017 at 03:26, Bas Nieuwenhuizen <bas at basnieuwenhuizen.nl> wrote:
> v2: Create the descriptor BO with CPU access.
> Signed-off-by: Bas Nieuwenhuizen <basni at google.com>

Tom has pushed the spill to llvm 4.0 as well now, so with the checks
changed to 4.0

for the series:
Reviewed-by: Dave Airlie <airlied at redhat.com>

> ---
>  src/amd/vulkan/radv_device.c   | 186 ++++++++++++++++++++++++++++++++++++++++-
>  src/amd/vulkan/radv_pipeline.c |  11 +--
>  src/amd/vulkan/radv_private.h  |   8 ++
>  3 files changed, 199 insertions(+), 6 deletions(-)
>
> diff --git a/src/amd/vulkan/radv_device.c b/src/amd/vulkan/radv_device.c
> index 0df9e783af9..6d3a8100f74 100644
> --- a/src/amd/vulkan/radv_device.c
> +++ b/src/amd/vulkan/radv_device.c
> @@ -32,6 +32,7 @@
>  #include <fcntl.h>
>  #include <sys/stat.h>
>  #include "radv_private.h"
> +#include "radv_cs.h"
>  #include "util/strtod.h"
>
>  #include <xf86drm.h>
> @@ -752,6 +753,15 @@ radv_queue_finish(struct radv_queue *queue)
>  {
>         if (queue->hw_ctx)
>                 queue->device->ws->ctx_destroy(queue->hw_ctx);
> +
> +       if (queue->preamble_cs)
> +               queue->device->ws->cs_destroy(queue->preamble_cs);
> +       if (queue->descriptor_bo)
> +               queue->device->ws->buffer_destroy(queue->descriptor_bo);
> +       if (queue->scratch_bo)
> +               queue->device->ws->buffer_destroy(queue->scratch_bo);
> +       if (queue->compute_scratch_bo)
> +               queue->device->ws->buffer_destroy(queue->compute_scratch_bo);
>  }
>
>  VkResult radv_CreateDevice(
> @@ -1001,6 +1011,159 @@ static void radv_dump_trace(struct radv_device *device,
>         fclose(f);
>  }
>
> +static VkResult
> +radv_get_preamble_cs(struct radv_queue *queue,
> +                     uint32_t scratch_size,
> +                     uint32_t compute_scratch_size,
> +                     struct radeon_winsys_cs **preamble_cs)
> +{
> +       struct radeon_winsys_bo *scratch_bo = NULL;
> +       struct radeon_winsys_bo *descriptor_bo = NULL;
> +       struct radeon_winsys_bo *compute_scratch_bo = NULL;
> +       struct radeon_winsys_cs *cs = NULL;
> +
> +       if (!scratch_size && !compute_scratch_size) {
> +               *preamble_cs = NULL;
> +               return VK_SUCCESS;
> +       }
> +
> +       if (scratch_size <= queue->scratch_size &&
> +           compute_scratch_size <= queue->compute_scratch_size) {
> +               *preamble_cs = queue->preamble_cs;
> +               return VK_SUCCESS;
> +       }
> +
> +       if (scratch_size > queue->scratch_size) {
> +               scratch_bo = queue->device->ws->buffer_create(queue->device->ws,
> +                                                             scratch_size,
> +                                                             4096,
> +                                                             RADEON_DOMAIN_VRAM,
> +                                                             RADEON_FLAG_NO_CPU_ACCESS);
> +               if (!scratch_bo)
> +                       goto fail;
> +       } else
> +               scratch_bo = queue->scratch_bo;
> +
> +       if (compute_scratch_size > queue->compute_scratch_size) {
> +               compute_scratch_bo = queue->device->ws->buffer_create(queue->device->ws,
> +                                                                     compute_scratch_size,
> +                                                                     4096,
> +                                                                     RADEON_DOMAIN_VRAM,
> +                                                                     RADEON_FLAG_NO_CPU_ACCESS);
> +               if (!compute_scratch_bo)
> +                       goto fail;
> +
> +       } else
> +               compute_scratch_bo = queue->compute_scratch_bo;
> +
> +       if (scratch_bo != queue->scratch_bo) {
> +               descriptor_bo = queue->device->ws->buffer_create(queue->device->ws,
> +                                                                8,
> +                                                                4096,
> +                                                                RADEON_DOMAIN_VRAM,
> +                                                                RADEON_FLAG_CPU_ACCESS);
> +               if (!descriptor_bo)
> +                       goto fail;
> +       } else
> +               descriptor_bo = queue->descriptor_bo;
> +
> +       cs = queue->device->ws->cs_create(queue->device->ws,
> +                                         queue->queue_family_index ? RING_COMPUTE : RING_GFX);
> +       if (!cs)
> +               goto fail;
> +
> +
> +       if (scratch_bo)
> +               queue->device->ws->cs_add_buffer(cs, scratch_bo, 8);
> +
> +       if (descriptor_bo)
> +               queue->device->ws->cs_add_buffer(cs, descriptor_bo, 8);
> +
> +       if (descriptor_bo != queue->descriptor_bo) {
> +               uint64_t scratch_va = queue->device->ws->buffer_get_va(scratch_bo);
> +               uint32_t rsrc1 = S_008F04_BASE_ADDRESS_HI(scratch_va >> 32) |
> +                                S_008F04_SWIZZLE_ENABLE(1);
> +
> +               uint32_t *map = (uint32_t*)queue->device->ws->buffer_map(descriptor_bo);
> +
> +               map[0] = scratch_va;
> +               map[1] = rsrc1;
> +
> +               queue->device->ws->buffer_unmap(descriptor_bo);
> +       }
> +
> +       if (descriptor_bo) {
> +               uint32_t regs[] = {R_00B030_SPI_SHADER_USER_DATA_PS_0,
> +                                  R_00B130_SPI_SHADER_USER_DATA_VS_0,
> +                                  R_00B230_SPI_SHADER_USER_DATA_GS_0,
> +                                  R_00B330_SPI_SHADER_USER_DATA_ES_0,
> +                                  R_00B430_SPI_SHADER_USER_DATA_HS_0,
> +                                  R_00B530_SPI_SHADER_USER_DATA_LS_0};
> +
> +               uint64_t va = queue->device->ws->buffer_get_va(descriptor_bo);
> +
> +               for (int i = 0; i < ARRAY_SIZE(regs); ++i) {
> +                       radeon_set_sh_reg_seq(cs, regs[i], 2);
> +                       radeon_emit(cs, va);
> +                       radeon_emit(cs, va >> 32);
> +               }
> +       }
> +
> +       if (compute_scratch_bo) {
> +               uint64_t scratch_va = queue->device->ws->buffer_get_va(compute_scratch_bo);
> +               uint32_t rsrc1 = S_008F04_BASE_ADDRESS_HI(scratch_va >> 32) |
> +                                S_008F04_SWIZZLE_ENABLE(1);
> +
> +               queue->device->ws->cs_add_buffer(cs, compute_scratch_bo, 8);
> +
> +               radeon_set_sh_reg_seq(cs, R_00B900_COMPUTE_USER_DATA_0, 2);
> +               radeon_emit(cs, scratch_va);
> +               radeon_emit(cs, rsrc1);
> +       }
> +
> +       if (!queue->device->ws->cs_finalize(cs))
> +               goto fail;
> +
> +       if (queue->preamble_cs)
> +               queue->device->ws->cs_destroy(queue->preamble_cs);
> +
> +       queue->preamble_cs = cs;
> +
> +       if (scratch_bo != queue->scratch_bo) {
> +               if (queue->scratch_bo)
> +                       queue->device->ws->buffer_destroy(queue->scratch_bo);
> +               queue->scratch_bo = scratch_bo;
> +               queue->scratch_size = scratch_size;
> +       }
> +
> +       if (compute_scratch_bo != queue->compute_scratch_bo) {
> +               if (queue->compute_scratch_bo)
> +                       queue->device->ws->buffer_destroy(queue->compute_scratch_bo);
> +               queue->compute_scratch_bo = compute_scratch_bo;
> +               queue->compute_scratch_size = compute_scratch_size;
> +       }
> +
> +       if (descriptor_bo != queue->descriptor_bo) {
> +               if (queue->descriptor_bo)
> +                       queue->device->ws->buffer_destroy(queue->descriptor_bo);
> +
> +               queue->descriptor_bo = descriptor_bo;
> +       }
> +
> +       *preamble_cs = cs;
> +       return VK_SUCCESS;
> +fail:
> +       if (cs)
> +               queue->device->ws->cs_destroy(cs);
> +       if (descriptor_bo && descriptor_bo != queue->descriptor_bo)
> +               queue->device->ws->buffer_destroy(descriptor_bo);
> +       if (scratch_bo && scratch_bo != queue->scratch_bo)
> +               queue->device->ws->buffer_destroy(scratch_bo);
> +       if (compute_scratch_bo && compute_scratch_bo != queue->compute_scratch_bo)
> +               queue->device->ws->buffer_destroy(compute_scratch_bo);
> +       return VK_ERROR_OUT_OF_DEVICE_MEMORY;
> +}
> +
>  VkResult radv_QueueSubmit(
>         VkQueue                                     _queue,
>         uint32_t                                    submitCount,
> @@ -1013,6 +1176,27 @@ VkResult radv_QueueSubmit(
>         struct radeon_winsys_ctx *ctx = queue->hw_ctx;
>         int ret;
>         uint32_t max_cs_submission = queue->device->trace_bo ? 1 : UINT32_MAX;
> +       uint32_t scratch_size = 0;
> +       uint32_t compute_scratch_size = 0;
> +       struct radeon_winsys_cs *preamble_cs = NULL;
> +       VkResult result;
> +
> +       /* Do this first so failing to allocate scratch buffers can't result in
> +        * partially executed submissions. */
> +       for (uint32_t i = 0; i < submitCount; i++) {
> +               for (uint32_t j = 0; j < pSubmits[i].commandBufferCount; j++) {
> +                       RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer,
> +                                        pSubmits[i].pCommandBuffers[j]);
> +
> +                       scratch_size = MAX2(scratch_size, cmd_buffer->scratch_size_needed);
> +                       compute_scratch_size = MAX2(compute_scratch_size,
> +                                                   cmd_buffer->compute_scratch_size_needed);
> +               }
> +       }
> +
> +       result = radv_get_preamble_cs(queue, scratch_size, compute_scratch_size, &preamble_cs);
> +       if (result != VK_SUCCESS)
> +               return result;
>
>         for (uint32_t i = 0; i < submitCount; i++) {
>                 struct radeon_winsys_cs **cs_array;
> @@ -1045,7 +1229,7 @@ VkResult radv_QueueSubmit(
>                                 *queue->device->trace_id_ptr = 0;
>
>                         ret = queue->device->ws->cs_submit(ctx, queue->queue_idx, cs_array + j,
> -                                                       advance, NULL,
> +                                                       advance, preamble_cs,
>                                                         (struct radeon_winsys_sem **)pSubmits[i].pWaitSemaphores,
>                                                         b ? pSubmits[i].waitSemaphoreCount : 0,
>                                                         (struct radeon_winsys_sem **)pSubmits[i].pSignalSemaphores,
> diff --git a/src/amd/vulkan/radv_pipeline.c b/src/amd/vulkan/radv_pipeline.c
> index e332877e2ba..3f90fb3016f 100644
> --- a/src/amd/vulkan/radv_pipeline.c
> +++ b/src/amd/vulkan/radv_pipeline.c
> @@ -363,12 +363,13 @@ static void radv_fill_shader_variant(struct radv_device *device,
>                                      struct ac_shader_binary *binary,
>                                      gl_shader_stage stage)
>  {
> -       variant->code_size = binary->code_size;
>         bool scratch_enabled = variant->config.scratch_bytes_per_wave > 0;
>         unsigned vgpr_comp_cnt = 0;
>
> -       if (scratch_enabled)
> -               radv_finishme("shader scratch space");
> +       if (scratch_enabled && !device->llvm_supports_spill)
> +               radv_finishme("shader scratch support only available with LLVM 5.0");
> +
> +       variant->code_size = binary->code_size;
>
>         switch (stage) {
>         case MESA_SHADER_VERTEX:
> @@ -433,8 +434,8 @@ static struct radv_shader_variant *radv_shader_variant_create(struct radv_device
>         options.unsafe_math = !!(device->debug_flags & RADV_DEBUG_UNSAFE_MATH);
>         options.family = chip_family;
>         options.chip_class = device->physical_device->rad_info.chip_class;
> -       options.supports_spill = false;
> -       tm = ac_create_target_machine(chip_family, false);
> +       options.supports_spill = device->llvm_supports_spill;
> +       tm = ac_create_target_machine(chip_family, options.supports_spill);
>         ac_compile_nir_shader(tm, &binary, &variant->config,
>                               &variant->info, shader, &options, dump);
>         LLVMDisposeTargetMachine(tm);
> diff --git a/src/amd/vulkan/radv_private.h b/src/amd/vulkan/radv_private.h
> index 88e05595380..fac5b97153d 100644
> --- a/src/amd/vulkan/radv_private.h
> +++ b/src/amd/vulkan/radv_private.h
> @@ -467,6 +467,14 @@ struct radv_queue {
>         struct radeon_winsys_ctx                    *hw_ctx;
>         int queue_family_index;
>         int queue_idx;
> +
> +       uint32_t scratch_size;
> +       uint32_t compute_scratch_size;
> +
> +       struct radeon_winsys_bo *scratch_bo;
> +       struct radeon_winsys_bo *descriptor_bo;
> +       struct radeon_winsys_bo *compute_scratch_bo;
> +       struct radeon_winsys_cs *preamble_cs;
>  };
>
>  struct radv_device {
> --
> 2.11.0
>
> _______________________________________________
> 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