panfrost: Describe thread local storage sizing rules

Deeply nested powers-of-two, basically :-)

Signed-off-by: Alyssa Rosenzweig <[email protected]>

2 files changed, 80 insertions, 0 deletions

diff --git a/src/panfrost/encoder/meson.build b/src/panfrost/encoder/meson.build
index 007785769af..310772d59c5 100644
--- a/src/panfrost/encoder/meson.build
+++ b/src/panfrost/encoder/meson.build
@@ -24,6 +24,7 @@ libpanfrost_encoder_files = files(
 
   'pan_invocation.c',
   'pan_tiler.c',
+  'pan_scratch.c',
 )
 
 libpanfrost_encoder = static_library(
diff --git a/src/panfrost/encoder/pan_scratch.c b/src/panfrost/encoder/pan_scratch.c
new file mode 100644
index 00000000000..4a0561c7383
--- /dev/null
+++ b/src/panfrost/encoder/pan_scratch.c
@@ -0,0 +1,79 @@
+/*
+ * Copyright (C) 2019 Collabora, Ltd.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ * Authors:
+ *   Alyssa Rosenzweig <[email protected]>
+ */
+
+#include "util/u_math.h"
+#include "pan_encoder.h"
+
+/* Midgard has a small register file, so shaders with high register pressure
+ * need to spill from the register file onto the stack. In addition to
+ * spilling, it is desireable to allocate temporary arrays on the stack (for
+ * instance because the register file does not support indirect access but the
+ * stack does).
+ *
+ * The stack is located in "Thread Local Storage", sometimes abbreviated TLS in
+ * the kernel source code. Thread local storage is allocated per-thread,
+ * per-core, so threads executing concurrently do not interfere with each
+ * other's stacks. On modern kernels, we may query
+ * DRM_PANFROST_PARAM_THREAD_TLS_ALLOC for the number of threads per core we
+ * must allocate for, and DRM_PANFROST_PARAM_SHADER_PRESENT for a bitmask of
+ * shader cores (so take a popcount of that mask for the number of shader
+ * cores). On older kernels that do not support querying these values,
+ * following kbase, we may use the worst-case value of 1024 threads for
+ * THREAD_TLS_ALLOC, and the worst-case value of 16 cores for Midgard per the
+ * "shader core count" column of the implementations table in
+ * https://en.wikipedia.org/wiki/Mali_%28GPU% [citation needed]
+ *
+ * Within a particular thread, there is stack allocated. If it is present, its
+ * size is a power-of-two, and it is at least 256 bytes. Stack is allocated
+ * with the framebuffer descriptor used for all shaders within a frame (note
+ * that they don't execute concurrently so it's fine). So, consider the maximum
+ * stack size used by any shader within a job, and then compute (where npot
+ * denotes the next power of two):
+ *
+ *      allocated = npot(max(size, 256)) * (# of threads/core) * (# of cores)
+ *
+ * The size of Thread Local Storage is signaled to the GPU in a dedicated
+ * log_stack_size field. Since stack sizes are powers of two, it follows that
+ * stack_size is logarithmic. Consider some sample values:
+ *
+ *      stack size | log_stack_size
+ *      ---------------------------
+ *             256 | 4
+ *             512 | 5
+ *            1024 | 6
+ *
+ *  Noting that log2(256) = 8, we have the relation:
+ *
+ *      stack_size <= 2^(log_stack_size + 4)
+ *
+ *  Given the constraints about powers-of-two and the minimum of 256, we thus
+ *  derive a formula for log_stack_size in terms of stack size (s):
+ *
+ *      log_stack_size = ceil(log2(max(s, 256))) - 4
+ *
+ * There are other valid characterisations of this formula, of course, but this
+ * is computationally simple, so good enough for our purposes.
+ */