From 8a9e67b8df9836408270a4bc3660ec45b622ae56 Mon Sep 17 00:00:00 2001 From: Chris Wilson Date: Thu, 10 Feb 2011 00:25:17 +0000 Subject: intel: Buffered upload Rather than performing lots of little writes to update the common bo upon each update, write those into a static buffer and flush that when full (or at the end of the batch). Doing so gives a dramatic performance improvement over and above using mmaped access. Signed-off-by: Chris Wilson --- src/mesa/drivers/dri/intel/intel_batchbuffer.c | 7 +- src/mesa/drivers/dri/intel/intel_buffer_objects.c | 85 ++++++++++++++++++++--- src/mesa/drivers/dri/intel/intel_buffer_objects.h | 2 + src/mesa/drivers/dri/intel/intel_context.h | 3 + 4 files changed, 81 insertions(+), 16 deletions(-) (limited to 'src/mesa/drivers/dri/intel') diff --git a/src/mesa/drivers/dri/intel/intel_batchbuffer.c b/src/mesa/drivers/dri/intel/intel_batchbuffer.c index 28bf42e770..5564e3ea2c 100644 --- a/src/mesa/drivers/dri/intel/intel_batchbuffer.c +++ b/src/mesa/drivers/dri/intel/intel_batchbuffer.c @@ -27,6 +27,7 @@ #include "intel_context.h" #include "intel_batchbuffer.h" +#include "intel_buffer_objects.h" #include "intel_decode.h" #include "intel_reg.h" #include "intel_bufmgr.h" @@ -175,11 +176,7 @@ _intel_batchbuffer_flush(struct intel_batchbuffer *batch, const char *file, if (intel->vtbl.finish_batch) intel->vtbl.finish_batch(intel); - if (intel->upload.bo) { - drm_intel_bo_unreference(intel->upload.bo); - intel->upload.bo = NULL; - intel->upload.offset = 0; - } + intel_upload_finish(intel); /* Check that we didn't just wrap our batchbuffer at a bad time. */ assert(!intel->no_batch_wrap); diff --git a/src/mesa/drivers/dri/intel/intel_buffer_objects.c b/src/mesa/drivers/dri/intel/intel_buffer_objects.c index d80d99bca5..19c73e6ab3 100644 --- a/src/mesa/drivers/dri/intel/intel_buffer_objects.c +++ b/src/mesa/drivers/dri/intel/intel_buffer_objects.c @@ -573,14 +573,31 @@ intel_bufferobj_buffer(struct intel_context *intel, #define INTEL_UPLOAD_SIZE (64*1024) +void +intel_upload_finish(struct intel_context *intel) +{ + if (!intel->upload.bo) + return; + + if (intel->upload.buffer_len) { + drm_intel_bo_subdata(intel->upload.bo, + intel->upload.buffer_offset, + intel->upload.buffer_len, + intel->upload.buffer); + intel->upload.buffer_len = 0; + } + + drm_intel_bo_unreference(intel->upload.bo); + intel->upload.bo = NULL; +} + static void wrap_buffers(struct intel_context *intel, GLuint size) { + intel_upload_finish(intel); + if (size < INTEL_UPLOAD_SIZE) size = INTEL_UPLOAD_SIZE; - if (intel->upload.bo != NULL) - drm_intel_bo_unreference(intel->upload.bo); - intel->upload.bo = drm_intel_bo_alloc(intel->bufmgr, "upload", size, 0); intel->upload.offset = 0; } @@ -590,6 +607,8 @@ void intel_upload_data(struct intel_context *intel, drm_intel_bo **return_bo, GLuint *return_offset) { + GLuint asize = ALIGN(size, 64); + if (intel->upload.bo == NULL || intel->upload.offset + size > intel->upload.bo->size) { wrap_buffers(intel, size); @@ -599,11 +618,32 @@ void intel_upload_data(struct intel_context *intel, *return_bo = intel->upload.bo; *return_offset = intel->upload.offset; - drm_intel_bo_subdata(intel->upload.bo, - intel->upload.offset, - size, ptr); + if (intel->upload.buffer_len && + intel->upload.buffer_len + asize > sizeof(intel->upload.buffer)) + { + drm_intel_bo_subdata(intel->upload.bo, + intel->upload.buffer_offset, + intel->upload.buffer_len, + intel->upload.buffer); + intel->upload.buffer_len = 0; + } + + if (size < sizeof(intel->upload.buffer)) + { + if (intel->upload.buffer_len == 0) + intel->upload.buffer_offset = intel->upload.offset; - intel->upload.offset += ALIGN(size, 64); + memcpy(intel->upload.buffer + intel->upload.buffer_len, ptr, size); + intel->upload.buffer_len += asize; + } + else + { + drm_intel_bo_subdata(intel->upload.bo, + intel->upload.offset, + size, ptr); + } + + intel->upload.offset += asize; } void *intel_upload_map(struct intel_context *intel, @@ -611,6 +651,7 @@ void *intel_upload_map(struct intel_context *intel, drm_intel_bo **return_bo, GLuint *return_offset) { + GLuint asize = ALIGN(size, 64); char *ptr; if (intel->upload.bo == NULL || @@ -622,10 +663,32 @@ void *intel_upload_map(struct intel_context *intel, *return_bo = intel->upload.bo; *return_offset = intel->upload.offset; - drm_intel_gem_bo_map_gtt(intel->upload.bo); - ptr = intel->upload.bo->virtual; - ptr += intel->upload.offset; - intel->upload.offset += ALIGN(size, 64); + if (intel->upload.buffer_len && + intel->upload.buffer_len + asize > sizeof(intel->upload.buffer)) + { + drm_intel_bo_subdata(intel->upload.bo, + intel->upload.buffer_offset, + intel->upload.buffer_len, + intel->upload.buffer); + intel->upload.buffer_len = 0; + } + + if (size < sizeof(intel->upload.buffer)) + { + if (intel->upload.buffer_len == 0) + intel->upload.buffer_offset = intel->upload.offset; + + ptr = intel->upload.buffer + intel->upload.buffer_len; + intel->upload.buffer_len += asize; + } + else + { + drm_intel_gem_bo_map_gtt(intel->upload.bo); + ptr = intel->upload.bo->virtual; + ptr += intel->upload.offset; + } + + intel->upload.offset += asize; return ptr; } diff --git a/src/mesa/drivers/dri/intel/intel_buffer_objects.h b/src/mesa/drivers/dri/intel/intel_buffer_objects.h index 71f0c52fd8..4b66bf7d59 100644 --- a/src/mesa/drivers/dri/intel/intel_buffer_objects.h +++ b/src/mesa/drivers/dri/intel/intel_buffer_objects.h @@ -80,6 +80,8 @@ void *intel_upload_map(struct intel_context *intel, drm_intel_bo **return_bo, GLuint *return_offset); +void intel_upload_finish(struct intel_context *intel); + /* Hook the bufferobject implementation into mesa: */ void intelInitBufferObjectFuncs(struct dd_function_table *functions); diff --git a/src/mesa/drivers/dri/intel/intel_context.h b/src/mesa/drivers/dri/intel/intel_context.h index a1ed462e54..0ea273837d 100644 --- a/src/mesa/drivers/dri/intel/intel_context.h +++ b/src/mesa/drivers/dri/intel/intel_context.h @@ -190,6 +190,9 @@ struct intel_context struct { drm_intel_bo *bo; GLuint offset; + uint16_t buffer_len; + uint16_t buffer_offset; + char buffer[4096]; } upload; GLuint stats_wm; -- cgit v1.2.3