fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-03 01:18:06 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

tu: Add a725 workaround dispatch at the start of each cmdbuf

Browse source 
Blob executes a special compute dispatch at the start of each
command buffers. We copy this dispatch as is. At this point
we don't know what this workaround is for.

Signed-off-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25888>
This commit is contained in:
Danylo Piliaiev 2023-11-10 18:25:01 +01:00 • committed by Marge Bot
parent 37f11ff1d4
commit 3cd6bb3e5d
6 changed files with 145 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

5
src/freedreno/common/freedreno_dev_info.h
View file

@ -201,6 +201,11 @@ struct fd_dev_info {
/* Whether there is CP_EVENT_WRITE7::WRITE_SAMPLE_COUNT */
bool has_event_write_sample_count;
/* Blob executes a special compute dispatch at the start of each
* command buffers. We copy this dispatch as is.
*/
bool cmdbuf_start_a725_quirk;
} a7xx;
};

4
src/freedreno/common/freedreno_devices.py
View file

@ -704,7 +704,9 @@ add_gpus([
)
))
a7xx_725 = A7XXProps()
a7xx_725 = A7XXProps(
cmdbuf_start_a725_quirk = True,
)
a7xx_730 = A7XXProps()

1
src/freedreno/registers/adreno/adreno_pm4.xml
View file

@ -1942,6 +1942,7 @@ opcode: CP_LOAD_STATE4 (30) (4 dwords)
<bitfield name="BV" pos="25" variants="THREAD_MODE" type="boolean"/>
<bitfield name="BR" pos="26" variants="THREAD_MODE" type="boolean"/>
<bitfield name="LPAC" pos="27" variants="THREAD_MODE" type="boolean"/>
<bitfield name="MODE" low="28" high="31" type="compare_mode" addvariant="yes"/>
</reg32>

9
src/freedreno/vulkan/tu_cmd_buffer.cc
View file

@ -1274,6 +1274,15 @@ tu6_init_hw(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_4(0x3f0243f0), );
}
if (phys_dev->info->a7xx.cmdbuf_start_a725_quirk) {
tu_cs_reserve(cs, 3 + 4);
tu_cs_emit_pkt7(cs, CP_COND_REG_EXEC, 2);
tu_cs_emit(cs, CP_COND_REG_EXEC_0_MODE(THREAD_MODE) |
CP_COND_REG_EXEC_0_BR | CP_COND_REG_EXEC_0_LPAC);
tu_cs_emit(cs, RENDER_MODE_CP_COND_REG_EXEC_1_DWORDS(4));
tu_cs_emit_ib(cs, dev->cmdbuf_start_a725_quirk_entry);
}
tu_cs_sanity_check(cs);
}

124
src/freedreno/vulkan/tu_device.cc
View file

@ -2016,6 +2016,112 @@ tu_init_dbg_reg_stomper(struct tu_device *device)
device->dbg_renderpass_stomp_cs = rp_cs;
}
/* It is unknown what this workaround is for and what it fixes. */
static VkResult
tu_init_cmdbuf_start_a725_quirk(struct tu_device *device)
{
struct tu_cs *cs;
if (!(device->cmdbuf_start_a725_quirk_cs =
(struct tu_cs *) calloc(1, sizeof(struct tu_cs)))) {
return vk_startup_errorf(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY,
"OOM");
}
if (!(device->cmdbuf_start_a725_quirk_entry =
(struct tu_cs_entry *) calloc(1, sizeof(struct tu_cs_entry)))) {
free(device->cmdbuf_start_a725_quirk_cs);
return vk_startup_errorf(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY,
"OOM");
}
cs = device->cmdbuf_start_a725_quirk_cs;
tu_cs_init(cs, device, TU_CS_MODE_SUB_STREAM, 57, "a725 workaround cs");
struct tu_cs shader_cs;
tu_cs_begin_sub_stream(cs, 10, &shader_cs);
uint32_t raw_shader[] = {
0x00040000, 0x40600000, // mul.f hr0.x, hr0.x, hr1.x
0x00050001, 0x40600001, // mul.f hr0.y, hr0.y, hr1.y
0x00060002, 0x40600002, // mul.f hr0.z, hr0.z, hr1.z
0x00070003, 0x40600003, // mul.f hr0.w, hr0.w, hr1.w
0x00000000, 0x03000000, // end
};
tu_cs_emit_array(&shader_cs, raw_shader, ARRAY_SIZE(raw_shader));
struct tu_cs_entry shader_entry = tu_cs_end_sub_stream(cs, &shader_cs);
uint64_t shader_iova = shader_entry.bo->iova + shader_entry.offset;
struct tu_cs sub_cs;
tu_cs_begin_sub_stream(cs, 47, &sub_cs);
tu_cs_emit_regs(&sub_cs, HLSQ_INVALIDATE_CMD(A7XX,
.vs_state = true, .hs_state = true, .ds_state = true,
.gs_state = true, .fs_state = true, .gfx_ibo = true,
.cs_bindless = 0xff, .gfx_bindless = 0xff));
tu_cs_emit_regs(&sub_cs, HLSQ_CS_CNTL(A7XX,
.constlen = 4,
.enabled = true));
tu_cs_emit_regs(&sub_cs, A6XX_SP_CS_CONFIG(.enabled = true));
tu_cs_emit_regs(&sub_cs, A6XX_SP_CS_CTRL_REG0(
.threadmode = MULTI,
.threadsize = THREAD128,
.mergedregs = true));
tu_cs_emit_regs(&sub_cs, A6XX_SP_CS_UNKNOWN_A9B1(.shared_size = 1));
tu_cs_emit_regs(&sub_cs, HLSQ_CS_KERNEL_GROUP_X(A7XX, 1),
HLSQ_CS_KERNEL_GROUP_Y(A7XX, 1),
HLSQ_CS_KERNEL_GROUP_Z(A7XX, 1));
tu_cs_emit_regs(&sub_cs, A6XX_SP_CS_INSTRLEN(.sp_cs_instrlen = 1));
tu_cs_emit_regs(&sub_cs, A6XX_SP_CS_TEX_COUNT(0));
tu_cs_emit_regs(&sub_cs, A6XX_SP_CS_IBO_COUNT(0));
tu_cs_emit_regs(&sub_cs, A7XX_HLSQ_CS_CNTL_1(
.linearlocalidregid = regid(63, 0),
.threadsize = THREAD128,
.unk11 = true,
.unk22 = true,
.yalign = CS_YALIGN_1));
tu_cs_emit_regs(&sub_cs, A6XX_SP_CS_CNTL_0(
.wgidconstid = regid(51, 3),
.wgsizeconstid = regid(48, 0),
.wgoffsetconstid = regid(63, 0),
.localidregid = regid(63, 0)));
tu_cs_emit_regs(&sub_cs, SP_CS_CNTL_1(A7XX,
.linearlocalidregid = regid(63, 0),
.threadsize = THREAD128,
.unk15 = true));
tu_cs_emit_regs(&sub_cs, A7XX_SP_CS_UNKNOWN_A9BE(0));
tu_cs_emit_regs(&sub_cs,
HLSQ_CS_NDRANGE_0(A7XX, .kerneldim = 3,
.localsizex = 255,
.localsizey = 1,
.localsizez = 1),
HLSQ_CS_NDRANGE_1(A7XX, .globalsize_x = 3072),
HLSQ_CS_NDRANGE_2(A7XX, .globaloff_x = 0),
HLSQ_CS_NDRANGE_3(A7XX, .globalsize_y = 1),
HLSQ_CS_NDRANGE_4(A7XX, .globaloff_y = 0),
HLSQ_CS_NDRANGE_5(A7XX, .globalsize_z = 1),
HLSQ_CS_NDRANGE_6(A7XX, .globaloff_z = 0));
tu_cs_emit_regs(&sub_cs, A7XX_HLSQ_CS_LOCAL_SIZE(
.localsizex = 255,
.localsizey = 0,
.localsizez = 0));
tu_cs_emit_pkt4(&sub_cs, REG_A6XX_SP_CS_OBJ_FIRST_EXEC_OFFSET, 3);
tu_cs_emit(&sub_cs, 0);
tu_cs_emit_qw(&sub_cs, shader_iova);
tu_cs_emit_pkt7(&sub_cs, CP_EXEC_CS, 4);
tu_cs_emit(&sub_cs, 0x00000000);
tu_cs_emit(&sub_cs, CP_EXEC_CS_1_NGROUPS_X(12));
tu_cs_emit(&sub_cs, CP_EXEC_CS_2_NGROUPS_Y(1));
tu_cs_emit(&sub_cs, CP_EXEC_CS_3_NGROUPS_Z(1));
*device->cmdbuf_start_a725_quirk_entry = tu_cs_end_sub_stream(cs, &sub_cs);
return VK_SUCCESS;
}
VKAPI_ATTR VkResult VKAPI_CALL
tu_CreateDevice(VkPhysicalDevice physicalDevice,
const VkDeviceCreateInfo *pCreateInfo,
@ -2315,6 +2421,12 @@ tu_CreateDevice(VkPhysicalDevice physicalDevice,
}
}
if (physical_device->info->a7xx.cmdbuf_start_a725_quirk) {
result = tu_init_cmdbuf_start_a725_quirk(device);
if (result != VK_SUCCESS)
goto fail_a725_workaround;
}
tu_init_dbg_reg_stomper(device);
/* Initialize a condition variable for timeline semaphore */
@ -2376,6 +2488,12 @@ tu_CreateDevice(VkPhysicalDevice physicalDevice,
return VK_SUCCESS;
fail_timeline_cond:
if (device->cmdbuf_start_a725_quirk_entry) {
free(device->cmdbuf_start_a725_quirk_entry);
tu_cs_finish(device->cmdbuf_start_a725_quirk_cs);
free(device->cmdbuf_start_a725_quirk_cs);
}
fail_a725_workaround:
fail_prepare_perfcntrs_pass_cs:
free(device->perfcntrs_pass_cs_entries);
tu_cs_finish(device->perfcntrs_pass_cs);
@ -2462,6 +2580,12 @@ tu_DestroyDevice(VkDevice _device, const VkAllocationCallbacks *pAllocator)
free(device->dbg_renderpass_stomp_cs);
}
if (device->cmdbuf_start_a725_quirk_entry) {
free(device->cmdbuf_start_a725_quirk_entry);
tu_cs_finish(device->cmdbuf_start_a725_quirk_cs);
free(device->cmdbuf_start_a725_quirk_cs);
}
tu_autotune_fini(&device->autotune, device);
tu_bo_suballocator_finish(&device->pipeline_suballoc);

3
src/freedreno/vulkan/tu_device.h
View file

@ -355,6 +355,9 @@ struct tu_device
struct tu_cs *perfcntrs_pass_cs;
struct tu_cs_entry *perfcntrs_pass_cs_entries;
struct tu_cs *cmdbuf_start_a725_quirk_cs;
struct tu_cs_entry *cmdbuf_start_a725_quirk_entry;
struct util_dynarray dynamic_rendering_pending;
VkCommandPool dynamic_rendering_pool;
uint32_t dynamic_rendering_fence;