radv: Handle command buffers that need scratch memory.
authorBas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Sun, 29 Jan 2017 12:53:05 +0000 (13:53 +0100)
committerBas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Mon, 30 Jan 2017 01:07:20 +0000 (02:07 +0100)
v2: Create the descriptor BO with CPU access.
Signed-off-by: Bas Nieuwenhuizen <basni@google.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
src/amd/vulkan/radv_device.c
src/amd/vulkan/radv_pipeline.c
src/amd/vulkan/radv_private.h

index da65511cf158c9a764f3c882f8aa51addfde7916..60714217f01c88d40b88604d898715e2862344a3 100644 (file)
@@ -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(
@@ -1000,6 +1010,159 @@ 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,
@@ -1012,6 +1175,27 @@ 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;
@@ -1044,7 +1228,7 @@ 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,
index e332877e2ba6b11633a11e44db32281533a8f875..f46987fdf6c6985c7e4ce31d07981879837d0424 100644 (file)
@@ -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 4.0");
+
+       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);
index 88e05595380008ddd2622cd2bae72b31b672bd88..fac5b97153df38b3f654cb6f2163a18645170e71 100644 (file)
@@ -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 {