leegao

// Operation: max(x[64,32] @ w[32,64], axis=1), Input in VMEM. Weights in HBM (need DMA).

   0x0   :  { %v_const_neg_inf = vmov -inf // Init accumulator for Max reduction
          ;;  %ptr_x_in      = inlined_call_operand.vmem [shape: f32[64,32]] // operand 0: input x
          ;;  %ptr_w_hbm     = inlined_call_operand.hbm  [shape: f32[32,64]] // operand 1: weights w (HBM)
          ;;  %ptr_out_max   = inlined_call_operand.vmem [shape: f32[64]]    // operand 2: output
          ;;  %ptr_out_full  = inlined_call_operand.vmem [shape: f32[64,64]] // operand 3: scratch }
   0x1   :  { %6 = vst [vmem:[#allocation1] sm:$0xff] /*vst_source=*/%v_const_neg_inf } // for another kernel, #allocation1 is an addr of -\infty

   // First Phase: get w out of HBM into vmem (the mxu and vpu can only use vmem)

leegao

!rm -rf compiler_dump
!rm compiler_dump.zip

import os

# # Create dump directories
DUMP_ROOT = "compiler_dump/"
HLO_DUMP_PATH = os.path.join(DUMP_ROOT, "hlo")
LLO_DUMP_PATH = os.path.join(DUMP_ROOT, "llo")

leegao

!rm -rf compiler_dump
!rm compiler_dump.zip

import os

# # Create dump directories
DUMP_ROOT = "compiler_dump/"
HLO_DUMP_PATH = os.path.join(DUMP_ROOT, "hlo")
LLO_DUMP_PATH = os.path.join(DUMP_ROOT, "llo")

leegao

Suppose little Gauss lived in the modern age. Little Gauss’ teacher wanted to surf the internet, so he assigned all of his students the following integral to evaluate:

$$ \int_0^1 x^{100} e^{x-1} dx $$

Being the clever alter-ego of the boy who immediately saw $\sum^n_k k = {n+1 \choose 2}$, little Gauss constructed a sequence

$$ s_k = \int_0^1 x^k e^{x-1} dx $$

and with a little bit of clever manipulation (integration by parts), he found that, using $u_k = x^k$, $dv = e^{x-1}dx$, $v = e^{x-1}$

leegao

Caching ASTC Parameters to Disk

Intro

Here's a negative result around compressing and caching ASTC parameters to reduce transfer overhead.

Taking an Adreno 650 as our reference:

1. GPU - 1.2 TFlops (1200 GFlops)
2. CPU SIMD int8 - 18 GFlops

leegao

[REQUIRED] depthBiasClamp / CORE10 (feature)
  FL=9.0 - Required for D3D9 and D3D11. Allows clamping the depth bias value.
   Qcom Turnip
    + unsupported (0/53): []
    +   supported (53/53): ['23.0-26.708 (52/52)']
   Qcom Proprietary
    + unsupported (0/163): []
    +   supported (163/163): ['512.502-512.826 (162/162)']
   Mali Proprietary
    + unsupported (0/67): []

leegao

CreateFramebuffer
  in: device: VkDevice (handle) = 0xb4000075a4e69010
  in: pCreateInfo: VkFramebufferCreateInfo*
    .flags: VkFramebufferCreateFlags = 0x0
    .renderPass: VkRenderPass (handle) = 0xb400007454e8d0d0
    .attachmentCount: uint32_t = 0x1
    .pAttachments[0]: VkImageView* = 0xb40000746501b800
    .width: uint32_t = 0x500
    .height: uint32_t = 0x2d0
    .layers: uint32_t = 0x1

leegao

From AI Studio analysis of https://github.com/google/angle/blob/6a04a50f98cac71b25464d10289ce7a013841caf/src/libANGLE/renderer/vulkan/vk_renderer.cpp#L4879

1. ARM / Mali

These workarounds apply to GPUs designed by ARM (Mali), found in chipsets like Samsung Exynos, Google Tensor, and MediaTek Dimensity.

Feature / Workaround	mFeatures Flag	Condition / Driver Version	Reason & Impact
Protected Memory Restriction	supportsProtectedMemory	Blocked if: isARM && !pipelineProtectedAccess	Bug: On older ARM platforms, enabling VK_KHR_protected_memory causes excessive, unnecessary load/store unit activity. Workaround: Only enabled on ARM if the newer VK_EXT_pipeline_protected_access extension is also present, indicating a fixed driver. (b/208458772)
Mixed Load Op Restriction	disallowMixedDepthStencilLoadOpNoneAndLoad	Enabled if: isARM && driverVersion < r38.1.0	Bug: ARM drivers older than r38p1 are bug

leegao

[REQUIRED] robustBufferAccess / CORE10 (feature)
  FL=9.1 - Always enabled if supported by Vulkan. Used for robustness and constant buffer range checks.
   ImgTec
    + unsupported (0/42): []
    +   supported (42/42): ['0.1017-139.3 (41/41)']
   Mali Proprietary
    + unsupported (0/66): []
    +   supported (66/66): ['25.1-50.0 (65/65)']
   Qcom Proprietary
    + unsupported (0/157): []

leegao

[
  {
    "name": "robustBufferAccess",
    "type": "feature",
    "extension": "CORE10",
    "required": true,
    "feature_level": "9.1",
    "notes": "Always enabled if supported by Vulkan. Used for robustness and constant buffer range checks.",
    "supported_driver_versions": {
      "Qcom Proprietary": [