#include "genxml/gen_macros.h"
#include "genxml/genX_pack.h"
+/* We reserve GPR 14 and 15 for conditional rendering */
+#define GEN_MI_BUILDER_NUM_ALLOC_GPRS 14
+#define __gen_get_batch_dwords anv_batch_emit_dwords
+#define __gen_address_offset anv_address_add
+#include "common/gen_mi_builder.h"
+
VkResult genX(CreateQueryPool)(
VkDevice _device,
const VkQueryPoolCreateInfo* pCreateInfo,
return status;
}
-static void
-emit_srm32(struct anv_batch *batch, struct anv_address addr, uint32_t reg)
-{
- anv_batch_emit(batch, GENX(MI_STORE_REGISTER_MEM), srm) {
- srm.MemoryAddress = addr;
- srm.RegisterAddress = reg;
- }
-}
-
-static void
-emit_srm64(struct anv_batch *batch, struct anv_address addr, uint32_t reg)
-{
- emit_srm32(batch, anv_address_add(addr, 0), reg + 0);
- emit_srm32(batch, anv_address_add(addr, 4), reg + 4);
-}
-
static void
emit_ps_depth_count(struct anv_cmd_buffer *cmd_buffer,
struct anv_address addr)
}
static void
-emit_query_availability(struct anv_cmd_buffer *cmd_buffer,
- struct anv_address addr)
+emit_query_mi_availability(struct gen_mi_builder *b,
+ struct anv_address addr,
+ bool available)
+{
+ gen_mi_store(b, gen_mi_mem64(addr), gen_mi_imm(available));
+}
+
+static void
+emit_query_pc_availability(struct anv_cmd_buffer *cmd_buffer,
+ struct anv_address addr,
+ bool available)
{
anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
pc.DestinationAddressType = DAT_PPGTT;
pc.PostSyncOperation = WriteImmediateData;
pc.Address = addr;
- pc.ImmediateData = 1;
+ pc.ImmediateData = available;
}
}
*/
static void
emit_zero_queries(struct anv_cmd_buffer *cmd_buffer,
- struct anv_query_pool *pool,
+ struct gen_mi_builder *b, struct anv_query_pool *pool,
uint32_t first_index, uint32_t num_queries)
{
- for (uint32_t i = 0; i < num_queries; i++) {
- struct anv_address slot_addr =
- anv_query_address(pool, first_index + i);
- genX(cmd_buffer_mi_memset)(cmd_buffer, anv_address_add(slot_addr, 8),
- 0, pool->stride - 8);
- emit_query_availability(cmd_buffer, slot_addr);
+ switch (pool->type) {
+ case VK_QUERY_TYPE_OCCLUSION:
+ case VK_QUERY_TYPE_TIMESTAMP:
+ /* These queries are written with a PIPE_CONTROL so clear them using the
+ * PIPE_CONTROL as well so we don't have to synchronize between 2 types
+ * of operations.
+ */
+ assert((pool->stride % 8) == 0);
+ for (uint32_t i = 0; i < num_queries; i++) {
+ struct anv_address slot_addr =
+ anv_query_address(pool, first_index + i);
+
+ for (uint32_t qword = 1; qword < (pool->stride / 8); qword++) {
+ emit_query_pc_availability(cmd_buffer,
+ anv_address_add(slot_addr, qword * 8),
+ false);
+ }
+ emit_query_pc_availability(cmd_buffer, slot_addr, true);
+ }
+ break;
+
+ case VK_QUERY_TYPE_PIPELINE_STATISTICS:
+ case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
+ for (uint32_t i = 0; i < num_queries; i++) {
+ struct anv_address slot_addr =
+ anv_query_address(pool, first_index + i);
+ gen_mi_memset(b, anv_address_add(slot_addr, 8), 0, pool->stride - 8);
+ emit_query_mi_availability(b, slot_addr, true);
+ }
+ break;
+
+ default:
+ unreachable("Unsupported query type");
}
}
ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
- for (uint32_t i = 0; i < queryCount; i++) {
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_DATA_IMM), sdm) {
- sdm.Address = anv_query_address(pool, firstQuery + i);
- sdm.ImmediateData = 0;
+ switch (pool->type) {
+ case VK_QUERY_TYPE_OCCLUSION:
+ case VK_QUERY_TYPE_TIMESTAMP:
+ for (uint32_t i = 0; i < queryCount; i++) {
+ emit_query_pc_availability(cmd_buffer,
+ anv_query_address(pool, firstQuery + i),
+ false);
}
+ break;
+
+ case VK_QUERY_TYPE_PIPELINE_STATISTICS:
+ case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT: {
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
+
+ for (uint32_t i = 0; i < queryCount; i++)
+ emit_query_mi_availability(&b, anv_query_address(pool, firstQuery + i), false);
+ break;
+ }
+
+ default:
+ unreachable("Unsupported query type");
}
}
};
static void
-emit_pipeline_stat(struct anv_cmd_buffer *cmd_buffer, uint32_t stat,
+emit_pipeline_stat(struct gen_mi_builder *b, uint32_t stat,
struct anv_address addr)
{
STATIC_ASSERT(ANV_PIPELINE_STATISTICS_MASK ==
(1 << ARRAY_SIZE(vk_pipeline_stat_to_reg)) - 1);
assert(stat < ARRAY_SIZE(vk_pipeline_stat_to_reg));
- emit_srm64(&cmd_buffer->batch, addr, vk_pipeline_stat_to_reg[stat]);
+ gen_mi_store(b, gen_mi_mem64(addr),
+ gen_mi_reg64(vk_pipeline_stat_to_reg[stat]));
}
static void
-emit_xfb_query(struct anv_cmd_buffer *cmd_buffer, uint32_t stream,
+emit_xfb_query(struct gen_mi_builder *b, uint32_t stream,
struct anv_address addr)
{
assert(stream < MAX_XFB_STREAMS);
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), lrm) {
- lrm.RegisterAddress = GENX(SO_NUM_PRIMS_WRITTEN0_num) + 0 + stream * 8;
- lrm.MemoryAddress = anv_address_add(addr, 0);
- }
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), lrm) {
- lrm.RegisterAddress = GENX(SO_NUM_PRIMS_WRITTEN0_num) + 4 + stream * 8;
- lrm.MemoryAddress = anv_address_add(addr, 4);
- }
-
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), lrm) {
- lrm.RegisterAddress = GENX(SO_PRIM_STORAGE_NEEDED0_num) + 0 + stream * 8;
- lrm.MemoryAddress = anv_address_add(addr, 16);
- }
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), lrm) {
- lrm.RegisterAddress = GENX(SO_PRIM_STORAGE_NEEDED0_num) + 4 + stream * 8;
- lrm.MemoryAddress = anv_address_add(addr, 20);
- }
+ gen_mi_store(b, gen_mi_mem64(anv_address_add(addr, 0)),
+ gen_mi_reg64(GENX(SO_NUM_PRIMS_WRITTEN0_num) + stream * 8));
+ gen_mi_store(b, gen_mi_mem64(anv_address_add(addr, 16)),
+ gen_mi_reg64(GENX(SO_PRIM_STORAGE_NEEDED0_num) + stream * 8));
}
void genX(CmdBeginQuery)(
ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
struct anv_address query_addr = anv_query_address(pool, query);
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
+
switch (pool->type) {
case VK_QUERY_TYPE_OCCLUSION:
emit_ps_depth_count(cmd_buffer, anv_address_add(query_addr, 8));
uint32_t offset = 8;
while (statistics) {
uint32_t stat = u_bit_scan(&statistics);
- emit_pipeline_stat(cmd_buffer, stat,
- anv_address_add(query_addr, offset));
+ emit_pipeline_stat(&b, stat, anv_address_add(query_addr, offset));
offset += 16;
}
break;
pc.CommandStreamerStallEnable = true;
pc.StallAtPixelScoreboard = true;
}
- emit_xfb_query(cmd_buffer, index, anv_address_add(query_addr, 8));
+ emit_xfb_query(&b, index, anv_address_add(query_addr, 8));
break;
default:
void genX(CmdEndQuery)(
VkCommandBuffer commandBuffer,
VkQueryPool queryPool,
- VkQueryControlFlags flags)
+ uint32_t query)
{
- genX(CmdEndQueryIndexedEXT)(commandBuffer, queryPool, flags, 0);
+ genX(CmdEndQueryIndexedEXT)(commandBuffer, queryPool, query, 0);
}
void genX(CmdEndQueryIndexedEXT)(
ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
struct anv_address query_addr = anv_query_address(pool, query);
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
+
switch (pool->type) {
case VK_QUERY_TYPE_OCCLUSION:
emit_ps_depth_count(cmd_buffer, anv_address_add(query_addr, 16));
- emit_query_availability(cmd_buffer, query_addr);
+ emit_query_pc_availability(cmd_buffer, query_addr, true);
break;
case VK_QUERY_TYPE_PIPELINE_STATISTICS: {
uint32_t offset = 16;
while (statistics) {
uint32_t stat = u_bit_scan(&statistics);
- emit_pipeline_stat(cmd_buffer, stat,
- anv_address_add(query_addr, offset));
+ emit_pipeline_stat(&b, stat, anv_address_add(query_addr, offset));
offset += 16;
}
- emit_query_availability(cmd_buffer, query_addr);
+ emit_query_mi_availability(&b, query_addr, true);
break;
}
pc.StallAtPixelScoreboard = true;
}
- emit_xfb_query(cmd_buffer, index, anv_address_add(query_addr, 16));
- emit_query_availability(cmd_buffer, query_addr);
+ emit_xfb_query(&b, index, anv_address_add(query_addr, 16));
+ emit_query_mi_availability(&b, query_addr, true);
break;
default:
const uint32_t num_queries =
util_bitcount(cmd_buffer->state.subpass->view_mask);
if (num_queries > 1)
- emit_zero_queries(cmd_buffer, pool, query + 1, num_queries - 1);
+ emit_zero_queries(cmd_buffer, &b, pool, query + 1, num_queries - 1);
}
}
assert(pool->type == VK_QUERY_TYPE_TIMESTAMP);
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
+
switch (pipelineStage) {
case VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT:
- emit_srm64(&cmd_buffer->batch, anv_address_add(query_addr, 8), TIMESTAMP);
+ gen_mi_store(&b, gen_mi_mem64(anv_address_add(query_addr, 8)),
+ gen_mi_reg64(TIMESTAMP));
break;
default:
break;
}
- emit_query_availability(cmd_buffer, query_addr);
+ emit_query_pc_availability(cmd_buffer, query_addr, true);
/* When multiview is active the spec requires that N consecutive query
* indices are used, where N is the number of active views in the subpass.
const uint32_t num_queries =
util_bitcount(cmd_buffer->state.subpass->view_mask);
if (num_queries > 1)
- emit_zero_queries(cmd_buffer, pool, query + 1, num_queries - 1);
+ emit_zero_queries(cmd_buffer, &b, pool, query + 1, num_queries - 1);
}
}
#if GEN_GEN > 7 || GEN_IS_HASWELL
-static uint32_t
-mi_alu(uint32_t opcode, uint32_t operand1, uint32_t operand2)
-{
- struct GENX(MI_MATH_ALU_INSTRUCTION) instr = {
- .ALUOpcode = opcode,
- .Operand1 = operand1,
- .Operand2 = operand2,
- };
-
- uint32_t dw;
- GENX(MI_MATH_ALU_INSTRUCTION_pack)(NULL, &dw, &instr);
-
- return dw;
-}
-
-#define CS_GPR(n) (0x2600 + (n) * 8)
-
static void
-emit_load_alu_reg_u64(struct anv_batch *batch, uint32_t reg,
- struct anv_address addr)
-{
- anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
- lrm.RegisterAddress = reg;
- lrm.MemoryAddress = anv_address_add(addr, 0);
- }
- anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
- lrm.RegisterAddress = reg + 4;
- lrm.MemoryAddress = anv_address_add(addr, 4);
- }
-}
-
-static void
-emit_load_alu_reg_imm32(struct anv_batch *batch, uint32_t reg, uint32_t imm)
-{
- anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_IMM), lri) {
- lri.RegisterOffset = reg;
- lri.DataDWord = imm;
- }
-}
-
-static void
-emit_load_alu_reg_imm64(struct anv_batch *batch, uint32_t reg, uint64_t imm)
-{
- emit_load_alu_reg_imm32(batch, reg, (uint32_t)imm);
- emit_load_alu_reg_imm32(batch, reg + 4, (uint32_t)(imm >> 32));
-}
-
-static void
-emit_load_alu_reg_reg32(struct anv_batch *batch, uint32_t src, uint32_t dst)
-{
- anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_REG), lrr) {
- lrr.SourceRegisterAddress = src;
- lrr.DestinationRegisterAddress = dst;
- }
-}
-
-/*
- * GPR0 = GPR0 & ((1ull << n) - 1);
- */
-static void
-keep_gpr0_lower_n_bits(struct anv_batch *batch, uint32_t n)
-{
- assert(n < 64);
- emit_load_alu_reg_imm64(batch, CS_GPR(1), (1ull << n) - 1);
-
- uint32_t *dw = anv_batch_emitn(batch, 5, GENX(MI_MATH));
- if (!dw) {
- anv_batch_set_error(batch, VK_ERROR_OUT_OF_HOST_MEMORY);
- return;
- }
-
- dw[1] = mi_alu(MI_ALU_LOAD, MI_ALU_SRCA, MI_ALU_REG0);
- dw[2] = mi_alu(MI_ALU_LOAD, MI_ALU_SRCB, MI_ALU_REG1);
- dw[3] = mi_alu(MI_ALU_AND, 0, 0);
- dw[4] = mi_alu(MI_ALU_STORE, MI_ALU_REG0, MI_ALU_ACCU);
-}
-
-/*
- * GPR0 = GPR0 << 30;
- */
-static void
-shl_gpr0_by_30_bits(struct anv_batch *batch)
-{
- /* First we mask 34 bits of GPR0 to prevent overflow */
- keep_gpr0_lower_n_bits(batch, 34);
-
- const uint32_t outer_count = 5;
- const uint32_t inner_count = 6;
- STATIC_ASSERT(outer_count * inner_count == 30);
- const uint32_t cmd_len = 1 + inner_count * 4;
-
- /* We'll emit 5 commands, each shifting GPR0 left by 6 bits, for a total of
- * 30 left shifts.
- */
- for (int o = 0; o < outer_count; o++) {
- /* Submit one MI_MATH to shift left by 6 bits */
- uint32_t *dw = anv_batch_emitn(batch, cmd_len, GENX(MI_MATH));
- if (!dw) {
- anv_batch_set_error(batch, VK_ERROR_OUT_OF_HOST_MEMORY);
- return;
- }
-
- dw++;
- for (int i = 0; i < inner_count; i++, dw += 4) {
- dw[0] = mi_alu(MI_ALU_LOAD, MI_ALU_SRCA, MI_ALU_REG0);
- dw[1] = mi_alu(MI_ALU_LOAD, MI_ALU_SRCB, MI_ALU_REG0);
- dw[2] = mi_alu(MI_ALU_ADD, 0, 0);
- dw[3] = mi_alu(MI_ALU_STORE, MI_ALU_REG0, MI_ALU_ACCU);
- }
- }
-}
-
-/*
- * GPR0 = GPR0 >> 2;
- *
- * Note that the upper 30 bits of GPR are lost!
- */
-static void
-shr_gpr0_by_2_bits(struct anv_batch *batch)
-{
- shl_gpr0_by_30_bits(batch);
- emit_load_alu_reg_reg32(batch, CS_GPR(0) + 4, CS_GPR(0));
- emit_load_alu_reg_imm32(batch, CS_GPR(0) + 4, 0);
-}
-
-static void
-gpu_write_query_result(struct anv_batch *batch,
+gpu_write_query_result(struct gen_mi_builder *b,
struct anv_address dst_addr,
VkQueryResultFlags flags,
- uint32_t value_index, uint32_t reg)
+ uint32_t value_index,
+ struct gen_mi_value query_result)
{
if (flags & VK_QUERY_RESULT_64_BIT) {
- emit_srm64(batch, anv_address_add(dst_addr, value_index * 8), reg);
+ struct anv_address res_addr = anv_address_add(dst_addr, value_index * 8);
+ gen_mi_store(b, gen_mi_mem64(res_addr), query_result);
} else {
- emit_srm32(batch, anv_address_add(dst_addr, value_index * 4), reg);
+ struct anv_address res_addr = anv_address_add(dst_addr, value_index * 4);
+ gen_mi_store(b, gen_mi_mem32(res_addr), query_result);
}
}
-static void
-compute_query_result(struct anv_batch *batch, uint32_t dst_reg,
- struct anv_address addr)
+static struct gen_mi_value
+compute_query_result(struct gen_mi_builder *b, struct anv_address addr)
{
- emit_load_alu_reg_u64(batch, CS_GPR(0), anv_address_add(addr, 0));
- emit_load_alu_reg_u64(batch, CS_GPR(1), anv_address_add(addr, 8));
-
- /* FIXME: We need to clamp the result for 32 bit. */
-
- uint32_t *dw = anv_batch_emitn(batch, 5, GENX(MI_MATH));
- if (!dw) {
- anv_batch_set_error(batch, VK_ERROR_OUT_OF_HOST_MEMORY);
- return;
- }
-
- dw[1] = mi_alu(MI_ALU_LOAD, MI_ALU_SRCA, MI_ALU_REG1);
- dw[2] = mi_alu(MI_ALU_LOAD, MI_ALU_SRCB, MI_ALU_REG0);
- dw[3] = mi_alu(MI_ALU_SUB, 0, 0);
- dw[4] = mi_alu(MI_ALU_STORE, dst_reg, MI_ALU_ACCU);
+ return gen_mi_isub(b, gen_mi_mem64(anv_address_add(addr, 8)),
+ gen_mi_mem64(anv_address_add(addr, 0)));
}
void genX(CmdCopyQueryPoolResults)(
ANV_FROM_HANDLE(anv_query_pool, pool, queryPool);
ANV_FROM_HANDLE(anv_buffer, buffer, destBuffer);
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
+ struct gen_mi_value result;
+
/* If render target writes are ongoing, request a render target cache flush
* to ensure proper ordering of the commands from the 3d pipe and the
* command streamer.
}
if ((flags & VK_QUERY_RESULT_WAIT_BIT) ||
- (cmd_buffer->state.pending_pipe_bits & ANV_PIPE_FLUSH_BITS)) {
+ (cmd_buffer->state.pending_pipe_bits & ANV_PIPE_FLUSH_BITS) ||
+ /* Occlusion & timestamp queries are written using a PIPE_CONTROL and
+ * because we're about to copy values from MI commands, we need to
+ * stall the command streamer to make sure the PIPE_CONTROL values have
+ * landed, otherwise we could see inconsistent values & availability.
+ *
+ * From the vulkan spec:
+ *
+ * "vkCmdCopyQueryPoolResults is guaranteed to see the effect of
+ * previous uses of vkCmdResetQueryPool in the same queue, without
+ * any additional synchronization."
+ */
+ pool->type == VK_QUERY_TYPE_OCCLUSION ||
+ pool->type == VK_QUERY_TYPE_TIMESTAMP) {
cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_CS_STALL_BIT;
genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
}
uint32_t idx = 0;
switch (pool->type) {
case VK_QUERY_TYPE_OCCLUSION:
- compute_query_result(&cmd_buffer->batch, MI_ALU_REG2,
- anv_address_add(query_addr, 8));
- gpu_write_query_result(&cmd_buffer->batch, dest_addr,
- flags, idx++, CS_GPR(2));
+ result = compute_query_result(&b, anv_address_add(query_addr, 8));
+ gpu_write_query_result(&b, dest_addr, flags, idx++, result);
break;
case VK_QUERY_TYPE_PIPELINE_STATISTICS: {
while (statistics) {
uint32_t stat = u_bit_scan(&statistics);
- compute_query_result(&cmd_buffer->batch, MI_ALU_REG0,
- anv_address_add(query_addr, idx * 16 + 8));
+ result = compute_query_result(&b, anv_address_add(query_addr,
+ idx * 16 + 8));
/* WaDividePSInvocationCountBy4:HSW,BDW */
if ((cmd_buffer->device->info.gen == 8 ||
cmd_buffer->device->info.is_haswell) &&
(1 << stat) == VK_QUERY_PIPELINE_STATISTIC_FRAGMENT_SHADER_INVOCATIONS_BIT) {
- shr_gpr0_by_2_bits(&cmd_buffer->batch);
+ result = gen_mi_ushr32_imm(&b, result, 2);
}
- gpu_write_query_result(&cmd_buffer->batch, dest_addr,
- flags, idx++, CS_GPR(0));
+ gpu_write_query_result(&b, dest_addr, flags, idx++, result);
}
assert(idx == util_bitcount(pool->pipeline_statistics));
break;
}
case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
- compute_query_result(&cmd_buffer->batch, MI_ALU_REG2,
- anv_address_add(query_addr, 8));
- gpu_write_query_result(&cmd_buffer->batch, dest_addr,
- flags, idx++, CS_GPR(2));
- compute_query_result(&cmd_buffer->batch, MI_ALU_REG2,
- anv_address_add(query_addr, 24));
- gpu_write_query_result(&cmd_buffer->batch, dest_addr,
- flags, idx++, CS_GPR(2));
+ result = compute_query_result(&b, anv_address_add(query_addr, 8));
+ gpu_write_query_result(&b, dest_addr, flags, idx++, result);
+ result = compute_query_result(&b, anv_address_add(query_addr, 24));
+ gpu_write_query_result(&b, dest_addr, flags, idx++, result);
break;
case VK_QUERY_TYPE_TIMESTAMP:
- emit_load_alu_reg_u64(&cmd_buffer->batch,
- CS_GPR(2), anv_address_add(query_addr, 8));
- gpu_write_query_result(&cmd_buffer->batch, dest_addr,
- flags, 0, CS_GPR(2));
+ result = gen_mi_mem64(anv_address_add(query_addr, 8));
+ gpu_write_query_result(&b, dest_addr, flags, 0, result);
break;
default:
}
if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
- emit_load_alu_reg_u64(&cmd_buffer->batch, CS_GPR(0), query_addr);
- gpu_write_query_result(&cmd_buffer->batch, dest_addr,
- flags, idx, CS_GPR(0));
+ gpu_write_query_result(&b, dest_addr, flags, idx,
+ gen_mi_mem64(query_addr));
}
dest_addr = anv_address_add(dest_addr, destStride);