From 81f30e2f509b4fcd79376ff02363aba831918ac6 Mon Sep 17 00:00:00 2001 From: Jordan Justen Date: Sat, 5 Mar 2016 00:54:54 -0800 Subject: [PATCH] anv/hsw: Move query code to genX file for Haswell This fixes many CTS cases, but will require an update to the kernel command parser register whitelist. (The CS GPRs and TIMESTAMP registers need to be whitelisted.) Signed-off-by: Jordan Justen --- src/intel/vulkan/gen8_cmd_buffer.c | 240 ---------------------------- src/intel/vulkan/genX_cmd_buffer.c | 244 +++++++++++++++++++++++++++++ 2 files changed, 244 insertions(+), 240 deletions(-) diff --git a/src/intel/vulkan/gen8_cmd_buffer.c b/src/intel/vulkan/gen8_cmd_buffer.c index 8972a8db6fc..8d8775fb01d 100644 --- a/src/intel/vulkan/gen8_cmd_buffer.c +++ b/src/intel/vulkan/gen8_cmd_buffer.c @@ -601,246 +601,6 @@ genX(cmd_buffer_flush_compute_state)(struct anv_cmd_buffer *cmd_buffer) cmd_buffer->state.compute_dirty = 0; } -static void -emit_ps_depth_count(struct anv_batch *batch, - struct anv_bo *bo, uint32_t offset) -{ - anv_batch_emit(batch, GENX(PIPE_CONTROL), - .DestinationAddressType = DAT_PPGTT, - .PostSyncOperation = WritePSDepthCount, - .DepthStallEnable = true, - .Address = { bo, offset }); -} - -static void -emit_query_availability(struct anv_batch *batch, - struct anv_bo *bo, uint32_t offset) -{ - anv_batch_emit(batch, GENX(PIPE_CONTROL), - .DestinationAddressType = DAT_PPGTT, - .PostSyncOperation = WriteImmediateData, - .Address = { bo, offset }, - .ImmediateData = 1); -} - -void genX(CmdBeginQuery)( - VkCommandBuffer commandBuffer, - VkQueryPool queryPool, - uint32_t query, - VkQueryControlFlags flags) -{ - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); - ANV_FROM_HANDLE(anv_query_pool, pool, queryPool); - - /* Workaround: When meta uses the pipeline with the VS disabled, it seems - * that the pipelining of the depth write breaks. What we see is that - * samples from the render pass clear leaks into the first query - * immediately after the clear. Doing a pipecontrol with a post-sync - * operation and DepthStallEnable seems to work around the issue. - */ - if (cmd_buffer->state.need_query_wa) { - cmd_buffer->state.need_query_wa = false; - anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), - .DepthCacheFlushEnable = true, - .DepthStallEnable = true); - } - - switch (pool->type) { - case VK_QUERY_TYPE_OCCLUSION: - emit_ps_depth_count(&cmd_buffer->batch, &pool->bo, - query * sizeof(struct anv_query_pool_slot)); - break; - - case VK_QUERY_TYPE_PIPELINE_STATISTICS: - default: - unreachable(""); - } -} - -void genX(CmdEndQuery)( - VkCommandBuffer commandBuffer, - VkQueryPool queryPool, - uint32_t query) -{ - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); - ANV_FROM_HANDLE(anv_query_pool, pool, queryPool); - - switch (pool->type) { - case VK_QUERY_TYPE_OCCLUSION: - emit_ps_depth_count(&cmd_buffer->batch, &pool->bo, - query * sizeof(struct anv_query_pool_slot) + 8); - - emit_query_availability(&cmd_buffer->batch, &pool->bo, - query * sizeof(struct anv_query_pool_slot) + 16); - break; - - case VK_QUERY_TYPE_PIPELINE_STATISTICS: - default: - unreachable(""); - } -} - -#define TIMESTAMP 0x2358 - -void genX(CmdWriteTimestamp)( - VkCommandBuffer commandBuffer, - VkPipelineStageFlagBits pipelineStage, - VkQueryPool queryPool, - uint32_t query) -{ - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); - ANV_FROM_HANDLE(anv_query_pool, pool, queryPool); - uint32_t offset = query * sizeof(struct anv_query_pool_slot); - - assert(pool->type == VK_QUERY_TYPE_TIMESTAMP); - - switch (pipelineStage) { - case VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT: - anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), - .RegisterAddress = TIMESTAMP, - .MemoryAddress = { &pool->bo, offset }); - anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), - .RegisterAddress = TIMESTAMP + 4, - .MemoryAddress = { &pool->bo, offset + 4 }); - break; - - default: - /* Everything else is bottom-of-pipe */ - anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), - .DestinationAddressType = DAT_PPGTT, - .PostSyncOperation = WriteTimestamp, - .Address = { &pool->bo, offset }); - break; - } - - emit_query_availability(&cmd_buffer->batch, &pool->bo, query + 16); -} - -#define alu_opcode(v) __gen_uint((v), 20, 31) -#define alu_operand1(v) __gen_uint((v), 10, 19) -#define alu_operand2(v) __gen_uint((v), 0, 9) -#define alu(opcode, operand1, operand2) \ - alu_opcode(opcode) | alu_operand1(operand1) | alu_operand2(operand2) - -#define OPCODE_NOOP 0x000 -#define OPCODE_LOAD 0x080 -#define OPCODE_LOADINV 0x480 -#define OPCODE_LOAD0 0x081 -#define OPCODE_LOAD1 0x481 -#define OPCODE_ADD 0x100 -#define OPCODE_SUB 0x101 -#define OPCODE_AND 0x102 -#define OPCODE_OR 0x103 -#define OPCODE_XOR 0x104 -#define OPCODE_STORE 0x180 -#define OPCODE_STOREINV 0x580 - -#define OPERAND_R0 0x00 -#define OPERAND_R1 0x01 -#define OPERAND_R2 0x02 -#define OPERAND_R3 0x03 -#define OPERAND_R4 0x04 -#define OPERAND_SRCA 0x20 -#define OPERAND_SRCB 0x21 -#define OPERAND_ACCU 0x31 -#define OPERAND_ZF 0x32 -#define OPERAND_CF 0x33 - -#define CS_GPR(n) (0x2600 + (n) * 8) - -static void -emit_load_alu_reg_u64(struct anv_batch *batch, uint32_t reg, - struct anv_bo *bo, uint32_t offset) -{ - anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_MEM), - .RegisterAddress = reg, - .MemoryAddress = { bo, offset }); - anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_MEM), - .RegisterAddress = reg + 4, - .MemoryAddress = { bo, offset + 4 }); -} - -static void -store_query_result(struct anv_batch *batch, uint32_t reg, - struct anv_bo *bo, uint32_t offset, VkQueryResultFlags flags) -{ - anv_batch_emit(batch, GENX(MI_STORE_REGISTER_MEM), - .RegisterAddress = reg, - .MemoryAddress = { bo, offset }); - - if (flags & VK_QUERY_RESULT_64_BIT) - anv_batch_emit(batch, GENX(MI_STORE_REGISTER_MEM), - .RegisterAddress = reg + 4, - .MemoryAddress = { bo, offset + 4 }); -} - -void genX(CmdCopyQueryPoolResults)( - VkCommandBuffer commandBuffer, - VkQueryPool queryPool, - uint32_t firstQuery, - uint32_t queryCount, - VkBuffer destBuffer, - VkDeviceSize destOffset, - VkDeviceSize destStride, - VkQueryResultFlags flags) -{ - ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); - ANV_FROM_HANDLE(anv_query_pool, pool, queryPool); - ANV_FROM_HANDLE(anv_buffer, buffer, destBuffer); - uint32_t slot_offset, dst_offset; - - if (flags & VK_QUERY_RESULT_WAIT_BIT) - anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), - .CommandStreamerStallEnable = true, - .StallAtPixelScoreboard = true); - - dst_offset = buffer->offset + destOffset; - for (uint32_t i = 0; i < queryCount; i++) { - - slot_offset = (firstQuery + i) * sizeof(struct anv_query_pool_slot); - switch (pool->type) { - case VK_QUERY_TYPE_OCCLUSION: - emit_load_alu_reg_u64(&cmd_buffer->batch, - CS_GPR(0), &pool->bo, slot_offset); - emit_load_alu_reg_u64(&cmd_buffer->batch, - CS_GPR(1), &pool->bo, slot_offset + 8); - - /* FIXME: We need to clamp the result for 32 bit. */ - - uint32_t *dw = anv_batch_emitn(&cmd_buffer->batch, 5, GENX(MI_MATH)); - dw[1] = alu(OPCODE_LOAD, OPERAND_SRCA, OPERAND_R1); - dw[2] = alu(OPCODE_LOAD, OPERAND_SRCB, OPERAND_R0); - dw[3] = alu(OPCODE_SUB, 0, 0); - dw[4] = alu(OPCODE_STORE, OPERAND_R2, OPERAND_ACCU); - break; - - case VK_QUERY_TYPE_TIMESTAMP: - emit_load_alu_reg_u64(&cmd_buffer->batch, - CS_GPR(2), &pool->bo, slot_offset); - break; - - default: - unreachable("unhandled query type"); - } - - store_query_result(&cmd_buffer->batch, - CS_GPR(2), buffer->bo, dst_offset, flags); - - if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) { - emit_load_alu_reg_u64(&cmd_buffer->batch, CS_GPR(0), - &pool->bo, slot_offset + 16); - if (flags & VK_QUERY_RESULT_64_BIT) - store_query_result(&cmd_buffer->batch, - CS_GPR(0), buffer->bo, dst_offset + 8, flags); - else - store_query_result(&cmd_buffer->batch, - CS_GPR(0), buffer->bo, dst_offset + 4, flags); - } - - dst_offset += destStride; - } -} - void genX(CmdSetEvent)( VkCommandBuffer commandBuffer, VkEvent _event, diff --git a/src/intel/vulkan/genX_cmd_buffer.c b/src/intel/vulkan/genX_cmd_buffer.c index a31ecc398e7..82959f3abf6 100644 --- a/src/intel/vulkan/genX_cmd_buffer.c +++ b/src/intel/vulkan/genX_cmd_buffer.c @@ -754,3 +754,247 @@ void genX(CmdEndRenderPass)( anv_cmd_buffer_resolve_subpass(cmd_buffer); } + +static void +emit_ps_depth_count(struct anv_batch *batch, + struct anv_bo *bo, uint32_t offset) +{ + anv_batch_emit(batch, GENX(PIPE_CONTROL), + .DestinationAddressType = DAT_PPGTT, + .PostSyncOperation = WritePSDepthCount, + .DepthStallEnable = true, + .Address = { bo, offset }); +} + +static void +emit_query_availability(struct anv_batch *batch, + struct anv_bo *bo, uint32_t offset) +{ + anv_batch_emit(batch, GENX(PIPE_CONTROL), + .DestinationAddressType = DAT_PPGTT, + .PostSyncOperation = WriteImmediateData, + .Address = { bo, offset }, + .ImmediateData = 1); +} + +void genX(CmdBeginQuery)( + VkCommandBuffer commandBuffer, + VkQueryPool queryPool, + uint32_t query, + VkQueryControlFlags flags) +{ + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); + ANV_FROM_HANDLE(anv_query_pool, pool, queryPool); + + /* Workaround: When meta uses the pipeline with the VS disabled, it seems + * that the pipelining of the depth write breaks. What we see is that + * samples from the render pass clear leaks into the first query + * immediately after the clear. Doing a pipecontrol with a post-sync + * operation and DepthStallEnable seems to work around the issue. + */ + if (cmd_buffer->state.need_query_wa) { + cmd_buffer->state.need_query_wa = false; + anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), + .DepthCacheFlushEnable = true, + .DepthStallEnable = true); + } + + switch (pool->type) { + case VK_QUERY_TYPE_OCCLUSION: + emit_ps_depth_count(&cmd_buffer->batch, &pool->bo, + query * sizeof(struct anv_query_pool_slot)); + break; + + case VK_QUERY_TYPE_PIPELINE_STATISTICS: + default: + unreachable(""); + } +} + +void genX(CmdEndQuery)( + VkCommandBuffer commandBuffer, + VkQueryPool queryPool, + uint32_t query) +{ + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); + ANV_FROM_HANDLE(anv_query_pool, pool, queryPool); + + switch (pool->type) { + case VK_QUERY_TYPE_OCCLUSION: + emit_ps_depth_count(&cmd_buffer->batch, &pool->bo, + query * sizeof(struct anv_query_pool_slot) + 8); + + emit_query_availability(&cmd_buffer->batch, &pool->bo, + query * sizeof(struct anv_query_pool_slot) + 16); + break; + + case VK_QUERY_TYPE_PIPELINE_STATISTICS: + default: + unreachable(""); + } +} + +#define TIMESTAMP 0x2358 + +void genX(CmdWriteTimestamp)( + VkCommandBuffer commandBuffer, + VkPipelineStageFlagBits pipelineStage, + VkQueryPool queryPool, + uint32_t query) +{ + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); + ANV_FROM_HANDLE(anv_query_pool, pool, queryPool); + uint32_t offset = query * sizeof(struct anv_query_pool_slot); + + assert(pool->type == VK_QUERY_TYPE_TIMESTAMP); + + switch (pipelineStage) { + case VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT: + anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), + .RegisterAddress = TIMESTAMP, + .MemoryAddress = { &pool->bo, offset }); + anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), + .RegisterAddress = TIMESTAMP + 4, + .MemoryAddress = { &pool->bo, offset + 4 }); + break; + + default: + /* Everything else is bottom-of-pipe */ + anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), + .DestinationAddressType = DAT_PPGTT, + .PostSyncOperation = WriteTimestamp, + .Address = { &pool->bo, offset }); + break; + } + + emit_query_availability(&cmd_buffer->batch, &pool->bo, query + 16); +} + +#if GEN_GEN > 7 || GEN_IS_HASWELL + +#define alu_opcode(v) __gen_uint((v), 20, 31) +#define alu_operand1(v) __gen_uint((v), 10, 19) +#define alu_operand2(v) __gen_uint((v), 0, 9) +#define alu(opcode, operand1, operand2) \ + alu_opcode(opcode) | alu_operand1(operand1) | alu_operand2(operand2) + +#define OPCODE_NOOP 0x000 +#define OPCODE_LOAD 0x080 +#define OPCODE_LOADINV 0x480 +#define OPCODE_LOAD0 0x081 +#define OPCODE_LOAD1 0x481 +#define OPCODE_ADD 0x100 +#define OPCODE_SUB 0x101 +#define OPCODE_AND 0x102 +#define OPCODE_OR 0x103 +#define OPCODE_XOR 0x104 +#define OPCODE_STORE 0x180 +#define OPCODE_STOREINV 0x580 + +#define OPERAND_R0 0x00 +#define OPERAND_R1 0x01 +#define OPERAND_R2 0x02 +#define OPERAND_R3 0x03 +#define OPERAND_R4 0x04 +#define OPERAND_SRCA 0x20 +#define OPERAND_SRCB 0x21 +#define OPERAND_ACCU 0x31 +#define OPERAND_ZF 0x32 +#define OPERAND_CF 0x33 + +#define CS_GPR(n) (0x2600 + (n) * 8) + +static void +emit_load_alu_reg_u64(struct anv_batch *batch, uint32_t reg, + struct anv_bo *bo, uint32_t offset) +{ + anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_MEM), + .RegisterAddress = reg, + .MemoryAddress = { bo, offset }); + anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_MEM), + .RegisterAddress = reg + 4, + .MemoryAddress = { bo, offset + 4 }); +} + +static void +store_query_result(struct anv_batch *batch, uint32_t reg, + struct anv_bo *bo, uint32_t offset, VkQueryResultFlags flags) +{ + anv_batch_emit(batch, GENX(MI_STORE_REGISTER_MEM), + .RegisterAddress = reg, + .MemoryAddress = { bo, offset }); + + if (flags & VK_QUERY_RESULT_64_BIT) + anv_batch_emit(batch, GENX(MI_STORE_REGISTER_MEM), + .RegisterAddress = reg + 4, + .MemoryAddress = { bo, offset + 4 }); +} + +void genX(CmdCopyQueryPoolResults)( + VkCommandBuffer commandBuffer, + VkQueryPool queryPool, + uint32_t firstQuery, + uint32_t queryCount, + VkBuffer destBuffer, + VkDeviceSize destOffset, + VkDeviceSize destStride, + VkQueryResultFlags flags) +{ + ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer); + ANV_FROM_HANDLE(anv_query_pool, pool, queryPool); + ANV_FROM_HANDLE(anv_buffer, buffer, destBuffer); + uint32_t slot_offset, dst_offset; + + if (flags & VK_QUERY_RESULT_WAIT_BIT) + anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), + .CommandStreamerStallEnable = true, + .StallAtPixelScoreboard = true); + + dst_offset = buffer->offset + destOffset; + for (uint32_t i = 0; i < queryCount; i++) { + + slot_offset = (firstQuery + i) * sizeof(struct anv_query_pool_slot); + switch (pool->type) { + case VK_QUERY_TYPE_OCCLUSION: + emit_load_alu_reg_u64(&cmd_buffer->batch, + CS_GPR(0), &pool->bo, slot_offset); + emit_load_alu_reg_u64(&cmd_buffer->batch, + CS_GPR(1), &pool->bo, slot_offset + 8); + + /* FIXME: We need to clamp the result for 32 bit. */ + + uint32_t *dw = anv_batch_emitn(&cmd_buffer->batch, 5, GENX(MI_MATH)); + dw[1] = alu(OPCODE_LOAD, OPERAND_SRCA, OPERAND_R1); + dw[2] = alu(OPCODE_LOAD, OPERAND_SRCB, OPERAND_R0); + dw[3] = alu(OPCODE_SUB, 0, 0); + dw[4] = alu(OPCODE_STORE, OPERAND_R2, OPERAND_ACCU); + break; + + case VK_QUERY_TYPE_TIMESTAMP: + emit_load_alu_reg_u64(&cmd_buffer->batch, + CS_GPR(2), &pool->bo, slot_offset); + break; + + default: + unreachable("unhandled query type"); + } + + store_query_result(&cmd_buffer->batch, + CS_GPR(2), buffer->bo, dst_offset, flags); + + if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) { + emit_load_alu_reg_u64(&cmd_buffer->batch, CS_GPR(0), + &pool->bo, slot_offset + 16); + if (flags & VK_QUERY_RESULT_64_BIT) + store_query_result(&cmd_buffer->batch, + CS_GPR(0), buffer->bo, dst_offset + 8, flags); + else + store_query_result(&cmd_buffer->batch, + CS_GPR(0), buffer->bo, dst_offset + 4, flags); + } + + dst_offset += destStride; + } +} + +#endif -- 2.30.2