On Wed, Nov 23, 2016 at 1:52 PM, Ilia Mirkin <imir...@alum.mit.edu> wrote:
> The strategy is to just keep n anv_query_pool_slot entries per query
> instead of one. The available bit is only valid in the last one.
>
> Signed-off-by: Ilia Mirkin <imir...@alum.mit.edu>
> ---
>
> This was lightly tested with this patch to the vulkan loader cube.c:
>
> http://pastie.org/private/fe92867hwhxu7fzjw9pvzq
>
> Flipping it to TYPE_OCCLUSION makes it show the same number of frag
> invocations.
>
> Without the giant stall, I didn't get any clipping primitives or frag
> invocations. I'm sure some of the bits are wholly unneccessary, but as I don't
> know anything about the architecture, I'm not inclined to guess which ones are
> and which ones aren't.
>
> No testing was done on the vkCmdCopyQueryPoolResults path.
>
> src/intel/vulkan/anv_device.c | 2 +-
> src/intel/vulkan/anv_private.h | 3 +
> src/intel/vulkan/anv_query.c | 87 +++++++++++++-----
> src/intel/vulkan/genX_cmd_buffer.c | 184
> +++++++++++++++++++++++++++++++------
> 4 files changed, 222 insertions(+), 54 deletions(-)
>
> diff --git a/src/intel/vulkan/anv_device.c b/src/intel/vulkan/anv_device.c
> index 7cce92c..fc8482c 100644
> --- a/src/intel/vulkan/anv_device.c
> +++ b/src/intel/vulkan/anv_device.c
> @@ -427,7 +427,7 @@ void anv_GetPhysicalDeviceFeatures(
> .textureCompressionASTC_LDR = pdevice->info.gen >= 9, /*
> FINISHME CHV */
> .textureCompressionBC = true,
> .occlusionQueryPrecise = true,
> - .pipelineStatisticsQuery = false,
> + .pipelineStatisticsQuery = true,
> .fragmentStoresAndAtomics = true,
> .shaderTessellationAndGeometryPointSize = true,
> .shaderImageGatherExtended = false,
> diff --git a/src/intel/vulkan/anv_private.h b/src/intel/vulkan/anv_private.h
> index 2fc543d..f033a8c 100644
> --- a/src/intel/vulkan/anv_private.h
> +++ b/src/intel/vulkan/anv_private.h
> @@ -1763,6 +1763,8 @@ struct anv_render_pass {
> struct anv_subpass subpasses[0];
> };
>
> +#define ANV_PIPELINE_STATISTICS_COUNT 11
> +
> struct anv_query_pool_slot {
> uint64_t begin;
> uint64_t end;
> @@ -1772,6 +1774,7 @@ struct anv_query_pool_slot {
> struct anv_query_pool {
> VkQueryType type;
> uint32_t slots;
> + uint32_t pipeline_statistics;
> struct anv_bo bo;
> };
>
> diff --git a/src/intel/vulkan/anv_query.c b/src/intel/vulkan/anv_query.c
> index 293257b..46e9d9a 100644
> --- a/src/intel/vulkan/anv_query.c
> +++ b/src/intel/vulkan/anv_query.c
> @@ -38,8 +38,8 @@ VkResult anv_CreateQueryPool(
> ANV_FROM_HANDLE(anv_device, device, _device);
> struct anv_query_pool *pool;
> VkResult result;
> - uint32_t slot_size;
> - uint64_t size;
> + uint32_t slot_size = sizeof(struct anv_query_pool_slot);
> + uint64_t size = pCreateInfo->queryCount * slot_size;
>
> assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO);
>
> @@ -48,12 +48,13 @@ VkResult anv_CreateQueryPool(
> case VK_QUERY_TYPE_TIMESTAMP:
> break;
> case VK_QUERY_TYPE_PIPELINE_STATISTICS:
> - return VK_ERROR_INCOMPATIBLE_DRIVER;
> + size *= _mesa_bitcount(pCreateInfo->pipelineStatistics);
> + break;
> default:
> assert(!"Invalid query type");
> + return VK_ERROR_INCOMPATIBLE_DRIVER;
> }
>
> - slot_size = sizeof(struct anv_query_pool_slot);
> pool = vk_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8,
> VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
> if (pool == NULL)
> @@ -61,8 +62,8 @@ VkResult anv_CreateQueryPool(
>
> pool->type = pCreateInfo->queryType;
> pool->slots = pCreateInfo->queryCount;
> + pool->pipeline_statistics = pCreateInfo->pipelineStatistics;
>
> - size = pCreateInfo->queryCount * slot_size;
> result = anv_bo_init_new(&pool->bo, device, size);
> if (result != VK_SUCCESS)
> goto fail;
> @@ -95,6 +96,27 @@ void anv_DestroyQueryPool(
> vk_free2(&device->alloc, pAllocator, pool);
> }
>
> +static void *
> +store_query_result(void *pData, VkQueryResultFlags flags,
> + uint64_t result, uint64_t available)
> +{
> + if (flags & VK_QUERY_RESULT_64_BIT) {
> + uint64_t *dst = pData;
> + *dst++ = result;
> + if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT)
> + *dst++ = available;
> + return dst;
> + } else {
> + uint32_t *dst = pData;
> + if (result > UINT32_MAX)
> + result = UINT32_MAX;
> + *dst++ = result;
> + if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT)
> + *dst++ = available;
> + return dst;
> + }
> +}
> +
> VkResult anv_GetQueryPoolResults(
> VkDevice _device,
> VkQueryPool queryPool,
> @@ -112,6 +134,7 @@ VkResult anv_GetQueryPoolResults(
> int ret;
>
> assert(pool->type == VK_QUERY_TYPE_OCCLUSION ||
> + pool->type == VK_QUERY_TYPE_PIPELINE_STATISTICS ||
> pool->type == VK_QUERY_TYPE_TIMESTAMP);
>
> if (pData == NULL)
> @@ -129,14 +152,35 @@ VkResult anv_GetQueryPoolResults(
> void *data_end = pData + dataSize;
> struct anv_query_pool_slot *slot = pool->bo.map;
>
> - for (uint32_t i = 0; i < queryCount; i++) {
> + for (uint32_t i = 0; i < queryCount && pData < data_end;
> + i++, pData += stride) {
> + if (pool->type == VK_QUERY_TYPE_PIPELINE_STATISTICS) {
> + VkQueryResultFlags f = flags &
> ~VK_QUERY_RESULT_WITH_AVAILABILITY_BIT;
> + void *pos = pData;
> + uint32_t queries = _mesa_bitcount(pool->pipeline_statistics);
> + struct anv_query_pool_slot *base = &slot[(firstQuery + i) *
> queries];
> +
> + for (uint32_t stat = 0; stat < ANV_PIPELINE_STATISTICS_COUNT;
> stat++) {
> + if (pool->pipeline_statistics & (1 << stat)) {
> + pos = store_query_result(pos, f, base->end - base->begin, 0);
> + base++;
> + }
> + }
> + if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
> + base--;
> + if (flags & VK_QUERY_RESULT_64_BIT)
> + *(uint64_t *)pos = base->available;
> + else
> + *(uint32_t *)pos = base->available;
> + }
> + continue;
> + }
> +
> switch (pool->type) {
> case VK_QUERY_TYPE_OCCLUSION: {
> result = slot[firstQuery + i].end - slot[firstQuery + i].begin;
> break;
> }
> - case VK_QUERY_TYPE_PIPELINE_STATISTICS:
> - unreachable("pipeline stats not supported");
> case VK_QUERY_TYPE_TIMESTAMP: {
> result = slot[firstQuery + i].begin;
> break;
> @@ -145,23 +189,7 @@ VkResult anv_GetQueryPoolResults(
> unreachable("invalid pool type");
> }
>
> - if (flags & VK_QUERY_RESULT_64_BIT) {
> - uint64_t *dst = pData;
> - dst[0] = result;
> - if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT)
> - dst[1] = slot[firstQuery + i].available;
> - } else {
> - uint32_t *dst = pData;
> - if (result > UINT32_MAX)
> - result = UINT32_MAX;
> - dst[0] = result;
> - if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT)
> - dst[1] = slot[firstQuery + i].available;
> - }
> -
> - pData += stride;
> - if (pData >= data_end)
> - break;
> + store_query_result(pData, flags, result, slot[firstQuery +
> i].available);
> }
>
> return VK_SUCCESS;
> @@ -183,6 +211,15 @@ void anv_CmdResetQueryPool(
> slot[firstQuery + i].available = 0;
> break;
> }
> + case VK_QUERY_TYPE_PIPELINE_STATISTICS: {
> + struct anv_query_pool_slot *slot = pool->bo.map;
> + uint32_t queries = _mesa_bitcount(pool->pipeline_statistics);
> +
> + slot = &slot[(firstQuery + i) * queries];
> + for (uint32_t j = 0; j < queries; j++)
> + slot[j].available = 0;
> + break;
> + }
> default:
> assert(!"Invalid query type");
> }
> diff --git a/src/intel/vulkan/genX_cmd_buffer.c
> b/src/intel/vulkan/genX_cmd_buffer.c
> index a965cd6..34c079c 100644
> --- a/src/intel/vulkan/genX_cmd_buffer.c
> +++ b/src/intel/vulkan/genX_cmd_buffer.c
> @@ -2272,6 +2272,55 @@ emit_query_availability(struct anv_cmd_buffer
> *cmd_buffer,
> }
> }
>
> +#define IA_VERTICES_COUNT 0x2310
> +#define IA_PRIMITIVES_COUNT 0x2318
> +#define VS_INVOCATION_COUNT 0x2320
> +#define HS_INVOCATION_COUNT 0x2300
> +#define DS_INVOCATION_COUNT 0x2308
> +#define GS_INVOCATION_COUNT 0x2328
> +#define GS_PRIMITIVES_COUNT 0x2330
> +#define CL_INVOCATION_COUNT 0x2338
> +#define CL_PRIMITIVES_COUNT 0x2340
> +#define PS_INVOCATION_COUNT 0x2348
> +#define CS_INVOCATION_COUNT 0x2290
> +
> +static const uint32_t PIPELINE_STAT_TO_REG[] = {
> + IA_VERTICES_COUNT,
> + IA_PRIMITIVES_COUNT,
> + VS_INVOCATION_COUNT,
> + GS_INVOCATION_COUNT,
> + GS_PRIMITIVES_COUNT,
> + CL_INVOCATION_COUNT,
> + CL_PRIMITIVES_COUNT,
> + PS_INVOCATION_COUNT,
> + HS_INVOCATION_COUNT,
> + DS_INVOCATION_COUNT,
> + CS_INVOCATION_COUNT
> +};
> +
> +static void
> +emit_pipeline_stat(struct anv_cmd_buffer *cmd_buffer, uint32_t stat,
> + struct anv_bo *bo, uint32_t offset) {
> + STATIC_ASSERT(ARRAY_SIZE(PIPELINE_STAT_TO_REG) ==
> + ANV_PIPELINE_STATISTICS_COUNT);
> +
> + uint32_t reg = PIPELINE_STAT_TO_REG[stat];
> +
> + /* TODO: Implement WaDividePSInvocationCountBy4:HSW,BDW
> + * NOTE: Experimentally, PS_INVOCATION_COUNT returns the same thing as
> + * the occlusion query on SKL.
> + */
> +
> + anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), lrm) {
> + lrm.RegisterAddress = reg,
> + lrm.MemoryAddress = (struct anv_address) { bo, offset };
> + }
> + anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), lrm) {
> + lrm.RegisterAddress = reg + 4,
> + lrm.MemoryAddress = (struct anv_address) { bo, offset + 4 };
> + }
> +}
> +
> void genX(CmdBeginQuery)(
> VkCommandBuffer commandBuffer,
> VkQueryPool queryPool,
> @@ -2301,7 +2350,31 @@ void genX(CmdBeginQuery)(
> query * sizeof(struct anv_query_pool_slot));
> break;
>
> - case VK_QUERY_TYPE_PIPELINE_STATISTICS:
> + case VK_QUERY_TYPE_PIPELINE_STATISTICS: {
> + uint32_t queries = _mesa_bitcount(pool->pipeline_statistics);
> + uint32_t q = 0;
> +
> + /* gen6_queryobj has this before emitting stats. Do we need it?
> + anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
> + pc.PostSyncOperation = NoWrite;
> + pc.RenderTargetCacheFlushEnable = true;
> + pc.InstructionCacheInvalidateEnable = true;
> + pc.DepthCacheFlushEnable = true;
> + pc.VFCacheInvalidationEnable = true;
> + pc.TextureCacheInvalidationEnable = true;
> + pc.CommandStreamerStallEnable = true;
> + }
> + */
Urgh, sorry - That was meant to terminate at the top of the comment
and do the PIPE_CONTROL emit. Forgot to fold my latest bit in. Same
for CmdEndQuery.
> + for (uint32_t stat = 0; stat < ANV_PIPELINE_STATISTICS_COUNT; stat++) {
> + if (pool->pipeline_statistics & (1 << stat)) {
> + emit_pipeline_stat(cmd_buffer, stat, &pool->bo,
> + (query * queries + q) *
> + sizeof(struct anv_query_pool_slot));
> + q++;
> + }
> + }
> + break;
> + }
> default:
> unreachable("");
> }
> @@ -2324,7 +2397,37 @@ void genX(CmdEndQuery)(
> query * sizeof(struct anv_query_pool_slot) +
> 16);
> break;
>
> - case VK_QUERY_TYPE_PIPELINE_STATISTICS:
> + case VK_QUERY_TYPE_PIPELINE_STATISTICS: {
> + uint32_t queries = _mesa_bitcount(pool->pipeline_statistics);
> + uint32_t q = 0;
> +
> + /* gen6_queryobj has this before emitting stats. Do we need it?
> + anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
> + pc.PostSyncOperation = NoWrite;
> + pc.RenderTargetCacheFlushEnable = true;
> + pc.InstructionCacheInvalidateEnable = true;
> + pc.DepthCacheFlushEnable = true;
> + pc.VFCacheInvalidationEnable = true;
> + pc.TextureCacheInvalidationEnable = true;
> + pc.CommandStreamerStallEnable = true;
> + }
> + */
> +
> + for (uint32_t stat = 0; stat < ANV_PIPELINE_STATISTICS_COUNT; stat++) {
> + if (pool->pipeline_statistics & (1 << stat)) {
> + emit_pipeline_stat(cmd_buffer, stat, &pool->bo,
> + (query * queries + q) *
> + sizeof(struct anv_query_pool_slot) + 8);
> + q++;
> + }
> + }
> +
> + emit_query_availability(cmd_buffer, &pool->bo,
> + (query * queries + q - 1) *
> + sizeof(struct anv_query_pool_slot) + 16);
> + break;
> + }
> +
> default:
> unreachable("");
> }
> @@ -2420,7 +2523,7 @@ emit_load_alu_reg_u64(struct anv_batch *batch, uint32_t
> reg,
> }
> }
>
> -static void
> +static uint32_t
> store_query_result(struct anv_batch *batch, uint32_t reg,
> struct anv_bo *bo, uint32_t offset, VkQueryResultFlags
> flags)
> {
> @@ -2434,7 +2537,27 @@ store_query_result(struct anv_batch *batch, uint32_t
> reg,
> srm.RegisterAddress = reg + 4;
> srm.MemoryAddress = (struct anv_address) { bo, offset + 4 };
> }
> +
> + return offset + 8;
> }
> +
> + return offset + 4;
> +}
> +
> +static void
> +compute_query_result(struct anv_batch *batch, struct anv_bo *bo,
> + uint32_t offset)
> +{
> + emit_load_alu_reg_u64(batch, CS_GPR(0), bo, offset);
> + emit_load_alu_reg_u64(batch, CS_GPR(1), bo, offset + 8);
> +
> + /* FIXME: We need to clamp the result for 32 bit. */
> +
> + uint32_t *dw = anv_batch_emitn(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);
> }
>
> void genX(CmdCopyQueryPoolResults)(
> @@ -2459,50 +2582,55 @@ void genX(CmdCopyQueryPoolResults)(
> }
> }
>
> - dst_offset = buffer->offset + destOffset;
> for (uint32_t i = 0; i < queryCount; i++) {
> -
> + dst_offset = buffer->offset + destOffset + destStride * 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);
> + compute_query_result(&cmd_buffer->batch, &pool->bo, slot_offset);
> + dst_offset = store_query_result(
> + &cmd_buffer->batch,
> + CS_GPR(2), buffer->bo, dst_offset, flags);
> break;
>
> case VK_QUERY_TYPE_TIMESTAMP:
> emit_load_alu_reg_u64(&cmd_buffer->batch,
> CS_GPR(2), &pool->bo, slot_offset);
> + dst_offset = store_query_result(
> + &cmd_buffer->batch,
> + CS_GPR(2), buffer->bo, dst_offset, flags);
> break;
>
> + case VK_QUERY_TYPE_PIPELINE_STATISTICS: {
> + uint32_t queries = _mesa_bitcount(pool->pipeline_statistics);
> +
> + slot_offset *= queries;
> + for (uint32_t slot = 0; slot < ANV_PIPELINE_STATISTICS_COUNT;
> slot++) {
> + if (pool->pipeline_statistics & (1 << slot)) {
> + compute_query_result(&cmd_buffer->batch, &pool->bo,
> slot_offset);
> + dst_offset = store_query_result(
> + &cmd_buffer->batch,
> + CS_GPR(2), buffer->bo, dst_offset, flags);
> + slot_offset += sizeof(struct anv_query_pool_slot);
> + }
> + }
> +
> + /* Get the slot offset to where it's supposed to be for the
> + * availability bit.
> + */
> + slot_offset -= sizeof(struct anv_query_pool_slot);
> + 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);
> + store_query_result(&cmd_buffer->batch,
> + CS_GPR(0), buffer->bo, dst_offset, flags);
> }
> -
> - dst_offset += destStride;
> }
> }
>
> --
> 2.7.3
>
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
